git.sesse.net Git - ffmpeg/blob - libavcodec/h264.c

   1 /*
   2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
   3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 /**
  23  * @file
  24  * H.264 / AVC / MPEG4 part10 codec.
  25  * @author Michael Niedermayer <michaelni@gmx.at>
  26  */
  27
  28 #include "libavutil/imgutils.h"
  29 #include "libavutil/opt.h"
  30 #include "internal.h"
  31 #include "dsputil.h"
  32 #include "avcodec.h"
  33 #include "mpegvideo.h"
  34 #include "h264.h"
  35 #include "h264data.h"
  36 #include "h264_mvpred.h"
  37 #include "golomb.h"
  38 #include "mathops.h"
  39 #include "rectangle.h"
  40 #include "thread.h"
  41 #include "vdpau_internal.h"
  42 #include "libavutil/avassert.h"
  43
  44 #include "cabac.h"
  45
  46 //#undef NDEBUG
  47 #include <assert.h>
  48
  49 static const uint8_t rem6[QP_MAX_NUM+1]={
  50 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
  51 };
  52
  53 static const uint8_t div6[QP_MAX_NUM+1]={
  54 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9,10,10,10,10,
  55 };
  56
  57 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
  58     PIX_FMT_DXVA2_VLD,
  59     PIX_FMT_VAAPI_VLD,
  60     PIX_FMT_VDA_VLD,
  61     PIX_FMT_YUVJ420P,
  62     PIX_FMT_NONE
  63 };
  64
  65 /**
  66  * Check if the top & left blocks are available if needed and
  67  * change the dc mode so it only uses the available blocks.
  68  */
  69 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
  70     MpegEncContext * const s = &h->s;
  71     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
  72     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
  73     int i;
  74
  75     if(!(h->top_samples_available&0x8000)){
  76         for(i=0; i<4; i++){
  77             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
  78             if(status<0){
  79                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
  80                 return -1;
  81             } else if(status){
  82                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
  83             }
  84         }
  85     }
  86
  87     if((h->left_samples_available&0x8888)!=0x8888){
  88         static const int mask[4]={0x8000,0x2000,0x80,0x20};
  89         for(i=0; i<4; i++){
  90             if(!(h->left_samples_available&mask[i])){
  91                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
  92                 if(status<0){
  93                     av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
  94                     return -1;
  95                 } else if(status){
  96                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
  97                 }
  98             }
  99         }
 100     }
 101
 102     return 0;
 103 } //FIXME cleanup like check_intra_pred_mode
 104
 105 static int check_intra_pred_mode(H264Context *h, int mode, int is_chroma){
 106     MpegEncContext * const s = &h->s;
 107     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
 108     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
 109
 110     if(mode > 6U) {
 111         av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
 112         return -1;
 113     }
 114
 115     if(!(h->top_samples_available&0x8000)){
 116         mode= top[ mode ];
 117         if(mode<0){
 118             av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
 119             return -1;
 120         }
 121     }
 122
 123     if((h->left_samples_available&0x8080) != 0x8080){
 124         mode= left[ mode ];
 125         if(is_chroma && (h->left_samples_available&0x8080)){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
 126             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
 127         }
 128         if(mode<0){
 129             av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
 130             return -1;
 131         }
 132     }
 133
 134     return mode;
 135 }
 136
 137 /**
 138  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 139  */
 140 int ff_h264_check_intra16x16_pred_mode(H264Context *h, int mode)
 141 {
 142     return check_intra_pred_mode(h, mode, 0);
 143 }
 144
 145 /**
 146  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 147  */
 148 int ff_h264_check_intra_chroma_pred_mode(H264Context *h, int mode)
 149 {
 150     return check_intra_pred_mode(h, mode, 1);
 151 }
 152
 153
 154 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
 155     int i, si, di;
 156     uint8_t *dst;
 157     int bufidx;
 158
 159 //    src[0]&0x80;                //forbidden bit
 160     h->nal_ref_idc= src[0]>>5;
 161     h->nal_unit_type= src[0]&0x1F;
 162
 163     src++; length--;
 164
 165 #if HAVE_FAST_UNALIGNED
 166 # if HAVE_FAST_64BIT
 167 #   define RS 7
 168     for(i=0; i+1<length; i+=9){
 169         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
 170 # else
 171 #   define RS 3
 172     for(i=0; i+1<length; i+=5){
 173         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
 174 # endif
 175             continue;
 176         if(i>0 && !src[i]) i--;
 177         while(src[i]) i++;
 178 #else
 179 #   define RS 0
 180     for(i=0; i+1<length; i+=2){
 181         if(src[i]) continue;
 182         if(i>0 && src[i-1]==0) i--;
 183 #endif
 184         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
 185             if(src[i+2]!=3){
 186                 /* startcode, so we must be past the end */
 187                 length=i;
 188             }
 189             break;
 190         }
 191         i-= RS;
 192     }
 193
 194     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
 195     si=h->rbsp_buffer_size[bufidx];
 196     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE+MAX_MBPAIR_SIZE);
 197     dst= h->rbsp_buffer[bufidx];
 198     if(si != h->rbsp_buffer_size[bufidx])
 199         memset(dst + length, 0, FF_INPUT_BUFFER_PADDING_SIZE+MAX_MBPAIR_SIZE);
 200
 201     if (dst == NULL){
 202         return NULL;
 203     }
 204
 205     if(i>=length-1){ //no escaped 0
 206         *dst_length= length;
 207         *consumed= length+1; //+1 for the header
 208         if(h->s.avctx->flags2 & CODEC_FLAG2_FAST){
 209             return src;
 210         }else{
 211             memcpy(dst, src, length);
 212             return dst;
 213         }
 214     }
 215
 216 //printf("decoding esc\n");
 217     memcpy(dst, src, i);
 218     si=di=i;
 219     while(si+2<length){
 220         //remove escapes (very rare 1:2^22)
 221         if(src[si+2]>3){
 222             dst[di++]= src[si++];
 223             dst[di++]= src[si++];
 224         }else if(src[si]==0 && src[si+1]==0){
 225             if(src[si+2]==3){ //escape
 226                 dst[di++]= 0;
 227                 dst[di++]= 0;
 228                 si+=3;
 229                 continue;
 230             }else //next start code
 231                 goto nsc;
 232         }
 233
 234         dst[di++]= src[si++];
 235     }
 236     while(si<length)
 237         dst[di++]= src[si++];
 238 nsc:
 239
 240     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 241
 242     *dst_length= di;
 243     *consumed= si + 1;//+1 for the header
 244 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
 245     return dst;
 246 }
 247
 248 /**
 249  * Identify the exact end of the bitstream
 250  * @return the length of the trailing, or 0 if damaged
 251  */
 252 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
 253     int v= *src;
 254     int r;
 255
 256     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
 257
 258     for(r=1; r<9; r++){
 259         if(v&1) return r;
 260         v>>=1;
 261     }
 262     return 0;
 263 }
 264
 265 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
 266                                  int y_offset, int list){
 267     int raw_my= h->mv_cache[list][ scan8[n] ][1];
 268     int filter_height= (raw_my&3) ? 2 : 0;
 269     int full_my= (raw_my>>2) + y_offset;
 270     int top = full_my - filter_height, bottom = full_my + height + filter_height;
 271
 272     return FFMAX(abs(top), bottom);
 273 }
 274
 275 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
 276                                int y_offset, int list0, int list1, int *nrefs){
 277     MpegEncContext * const s = &h->s;
 278     int my;
 279
 280     y_offset += 16*(s->mb_y >> MB_FIELD);
 281
 282     if(list0){
 283         int ref_n = h->ref_cache[0][ scan8[n] ];
 284         Picture *ref= &h->ref_list[0][ref_n];
 285
 286         // Error resilience puts the current picture in the ref list.
 287         // Don't try to wait on these as it will cause a deadlock.
 288         // Fields can wait on each other, though.
 289         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
 290            (ref->f.reference & 3) != s->picture_structure) {
 291             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
 292             if (refs[0][ref_n] < 0) nrefs[0] += 1;
 293             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
 294         }
 295     }
 296
 297     if(list1){
 298         int ref_n = h->ref_cache[1][ scan8[n] ];
 299         Picture *ref= &h->ref_list[1][ref_n];
 300
 301         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
 302            (ref->f.reference & 3) != s->picture_structure) {
 303             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
 304             if (refs[1][ref_n] < 0) nrefs[1] += 1;
 305             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
 306         }
 307     }
 308 }
 309
 310 /**
 311  * Wait until all reference frames are available for MC operations.
 312  *
 313  * @param h the H264 context
 314  */
 315 static void await_references(H264Context *h){
 316     MpegEncContext * const s = &h->s;
 317     const int mb_xy= h->mb_xy;
 318     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 319     int refs[2][48];
 320     int nrefs[2] = {0};
 321     int ref, list;
 322
 323     memset(refs, -1, sizeof(refs));
 324
 325     if(IS_16X16(mb_type)){
 326         get_lowest_part_y(h, refs, 0, 16, 0,
 327                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 328     }else if(IS_16X8(mb_type)){
 329         get_lowest_part_y(h, refs, 0, 8, 0,
 330                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 331         get_lowest_part_y(h, refs, 8, 8, 8,
 332                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 333     }else if(IS_8X16(mb_type)){
 334         get_lowest_part_y(h, refs, 0, 16, 0,
 335                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 336         get_lowest_part_y(h, refs, 4, 16, 0,
 337                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 338     }else{
 339         int i;
 340
 341         assert(IS_8X8(mb_type));
 342
 343         for(i=0; i<4; i++){
 344             const int sub_mb_type= h->sub_mb_type[i];
 345             const int n= 4*i;
 346             int y_offset= (i&2)<<2;
 347
 348             if(IS_SUB_8X8(sub_mb_type)){
 349                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 350                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 351             }else if(IS_SUB_8X4(sub_mb_type)){
 352                 get_lowest_part_y(h, refs, n  , 4, y_offset,
 353                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 354                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
 355                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 356             }else if(IS_SUB_4X8(sub_mb_type)){
 357                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 358                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 359                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
 360                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 361             }else{
 362                 int j;
 363                 assert(IS_SUB_4X4(sub_mb_type));
 364                 for(j=0; j<4; j++){
 365                     int sub_y_offset= y_offset + 2*(j&2);
 366                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
 367                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 368                 }
 369             }
 370         }
 371     }
 372
 373     for(list=h->list_count-1; list>=0; list--){
 374         for(ref=0; ref<48 && nrefs[list]; ref++){
 375             int row = refs[list][ref];
 376             if(row >= 0){
 377                 Picture *ref_pic = &h->ref_list[list][ref];
 378                 int ref_field = ref_pic->f.reference - 1;
 379                 int ref_field_picture = ref_pic->field_picture;
 380                 int pic_height = 16*s->mb_height >> ref_field_picture;
 381
 382                 row <<= MB_MBAFF;
 383                 nrefs[list]--;
 384
 385                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
 386                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
 387                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
 388                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
 389                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
 390                 }else if(FIELD_PICTURE){
 391                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
 392                 }else{
 393                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
 394                 }
 395             }
 396         }
 397     }
 398 }
 399
 400 #if 0
 401 /**
 402  * DCT transforms the 16 dc values.
 403  * @param qp quantization parameter ??? FIXME
 404  */
 405 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
 406 //    const int qmul= dequant_coeff[qp][0];
 407     int i;
 408     int temp[16]; //FIXME check if this is a good idea
 409     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
 410     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
 411
 412     for(i=0; i<4; i++){
 413         const int offset= y_offset[i];
 414         const int z0= block[offset+stride*0] + block[offset+stride*4];
 415         const int z1= block[offset+stride*0] - block[offset+stride*4];
 416         const int z2= block[offset+stride*1] - block[offset+stride*5];
 417         const int z3= block[offset+stride*1] + block[offset+stride*5];
 418
 419         temp[4*i+0]= z0+z3;
 420         temp[4*i+1]= z1+z2;
 421         temp[4*i+2]= z1-z2;
 422         temp[4*i+3]= z0-z3;
 423     }
 424
 425     for(i=0; i<4; i++){
 426         const int offset= x_offset[i];
 427         const int z0= temp[4*0+i] + temp[4*2+i];
 428         const int z1= temp[4*0+i] - temp[4*2+i];
 429         const int z2= temp[4*1+i] - temp[4*3+i];
 430         const int z3= temp[4*1+i] + temp[4*3+i];
 431
 432         block[stride*0 +offset]= (z0 + z3)>>1;
 433         block[stride*2 +offset]= (z1 + z2)>>1;
 434         block[stride*8 +offset]= (z1 - z2)>>1;
 435         block[stride*10+offset]= (z0 - z3)>>1;
 436     }
 437 }
 438 #endif
 439
 440 #undef xStride
 441 #undef stride
 442
 443 #if 0
 444 static void chroma_dc_dct_c(DCTELEM *block){
 445     const int stride= 16*2;
 446     const int xStride= 16;
 447     int a,b,c,d,e;
 448
 449     a= block[stride*0 + xStride*0];
 450     b= block[stride*0 + xStride*1];
 451     c= block[stride*1 + xStride*0];
 452     d= block[stride*1 + xStride*1];
 453
 454     e= a-b;
 455     a= a+b;
 456     b= c-d;
 457     c= c+d;
 458
 459     block[stride*0 + xStride*0]= (a+c);
 460     block[stride*0 + xStride*1]= (e+b);
 461     block[stride*1 + xStride*0]= (a-c);
 462     block[stride*1 + xStride*1]= (e-b);
 463 }
 464 #endif
 465
 466 static av_always_inline void
 467 mc_dir_part(H264Context *h, Picture *pic, int n, int square,
 468             int height, int delta, int list,
 469             uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 470             int src_x_offset, int src_y_offset,
 471             qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
 472             int pixel_shift, int chroma_idc)
 473 {
 474     MpegEncContext * const s = &h->s;
 475     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
 476     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
 477     const int luma_xy= (mx&3) + ((my&3)<<2);
 478     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
 479     uint8_t * src_y = pic->f.data[0] + offset;
 480     uint8_t * src_cb, * src_cr;
 481     int extra_width= h->emu_edge_width;
 482     int extra_height= h->emu_edge_height;
 483     int emu=0;
 484     const int full_mx= mx>>2;
 485     const int full_my= my>>2;
 486     const int pic_width  = 16*s->mb_width;
 487     const int pic_height = 16*s->mb_height >> MB_FIELD;
 488     int ysh;
 489
 490     if(mx&7) extra_width -= 3;
 491     if(my&7) extra_height -= 3;
 492
 493     if(   full_mx < 0-extra_width
 494        || full_my < 0-extra_height
 495        || full_mx + 16/*FIXME*/ > pic_width + extra_width
 496        || full_my + 16/*FIXME*/ > pic_height + extra_height){
 497         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_y - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 498                                 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 499             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 500         emu=1;
 501     }
 502
 503     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
 504     if(!square){
 505         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
 506     }
 507
 508     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
 509
 510     if(chroma_idc == 3 /* yuv444 */){
 511         src_cb = pic->f.data[1] + offset;
 512         if(emu){
 513             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 514                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 515             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 516         }
 517         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
 518         if(!square){
 519             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
 520         }
 521
 522         src_cr = pic->f.data[2] + offset;
 523         if(emu){
 524             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 525                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 526             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 527         }
 528         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
 529         if(!square){
 530             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
 531         }
 532         return;
 533     }
 534
 535     ysh = 3 - (chroma_idc == 2 /* yuv422 */);
 536     if(chroma_idc == 1 /* yuv420 */ && MB_FIELD){
 537         // chroma offset when predicting from a field of opposite parity
 538         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
 539         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
 540     }
 541
 542     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> ysh) * h->mb_uvlinesize;
 543     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> ysh) * h->mb_uvlinesize;
 544
 545     if(emu){
 546         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
 547                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 548                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 549             src_cb= s->edge_emu_buffer;
 550     }
 551     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
 552               mx&7, (my << (chroma_idc == 2 /* yuv422 */)) &7);
 553
 554     if(emu){
 555         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
 556                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 557                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 558             src_cr= s->edge_emu_buffer;
 559     }
 560     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
 561               mx&7, (my << (chroma_idc == 2 /* yuv422 */)) &7);
 562 }
 563
 564 static av_always_inline void
 565 mc_part_std(H264Context *h, int n, int square, int height, int delta,
 566             uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 567             int x_offset, int y_offset,
 568             qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 569             qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 570             int list0, int list1, int pixel_shift, int chroma_idc)
 571 {
 572     MpegEncContext * const s = &h->s;
 573     qpel_mc_func *qpix_op=  qpix_put;
 574     h264_chroma_mc_func chroma_op= chroma_put;
 575
 576     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 577     if (chroma_idc == 3 /* yuv444 */) {
 578         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 579         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 580     } else if (chroma_idc == 2 /* yuv422 */) {
 581         dest_cb += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
 582         dest_cr += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
 583     } else /* yuv420 */ {
 584         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 585         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 586     }
 587     x_offset += 8*s->mb_x;
 588     y_offset += 8*(s->mb_y >> MB_FIELD);
 589
 590     if(list0){
 591         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
 592         mc_dir_part(h, ref, n, square, height, delta, 0,
 593                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 594                            qpix_op, chroma_op, pixel_shift, chroma_idc);
 595
 596         qpix_op=  qpix_avg;
 597         chroma_op= chroma_avg;
 598     }
 599
 600     if(list1){
 601         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
 602         mc_dir_part(h, ref, n, square, height, delta, 1,
 603                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 604                            qpix_op, chroma_op, pixel_shift, chroma_idc);
 605     }
 606 }
 607
 608 static av_always_inline void
 609 mc_part_weighted(H264Context *h, int n, int square, int height, int delta,
 610                  uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 611                  int x_offset, int y_offset,
 612                  qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 613                  h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
 614                  h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
 615                  int list0, int list1, int pixel_shift, int chroma_idc){
 616     MpegEncContext * const s = &h->s;
 617     int chroma_height;
 618
 619     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 620     if (chroma_idc == 3 /* yuv444 */) {
 621         chroma_height = height;
 622         chroma_weight_avg = luma_weight_avg;
 623         chroma_weight_op = luma_weight_op;
 624         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 625         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 626     } else if (chroma_idc == 2 /* yuv422 */) {
 627         chroma_height = height;
 628         dest_cb += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
 629         dest_cr += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
 630     } else /* yuv420 */ {
 631         chroma_height = height >> 1;
 632         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 633         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 634     }
 635     x_offset += 8*s->mb_x;
 636     y_offset += 8*(s->mb_y >> MB_FIELD);
 637
 638     if(list0 && list1){
 639         /* don't optimize for luma-only case, since B-frames usually
 640          * use implicit weights => chroma too. */
 641         uint8_t *tmp_cb = s->obmc_scratchpad;
 642         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
 643         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
 644         int refn0 = h->ref_cache[0][ scan8[n] ];
 645         int refn1 = h->ref_cache[1][ scan8[n] ];
 646
 647         mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
 648                     dest_y, dest_cb, dest_cr,
 649                     x_offset, y_offset, qpix_put, chroma_put,
 650                     pixel_shift, chroma_idc);
 651         mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
 652                     tmp_y, tmp_cb, tmp_cr,
 653                     x_offset, y_offset, qpix_put, chroma_put,
 654                     pixel_shift, chroma_idc);
 655
 656         if(h->use_weight == 2){
 657             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
 658             int weight1 = 64 - weight0;
 659             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize,
 660                               height,        5, weight0, weight1, 0);
 661             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
 662                               chroma_height, 5, weight0, weight1, 0);
 663             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
 664                               chroma_height, 5, weight0, weight1, 0);
 665         }else{
 666             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height, h->luma_log2_weight_denom,
 667                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
 668                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
 669             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
 670                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
 671                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
 672             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
 673                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
 674                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
 675         }
 676     }else{
 677         int list = list1 ? 1 : 0;
 678         int refn = h->ref_cache[list][ scan8[n] ];
 679         Picture *ref= &h->ref_list[list][refn];
 680         mc_dir_part(h, ref, n, square, height, delta, list,
 681                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 682                     qpix_put, chroma_put, pixel_shift, chroma_idc);
 683
 684         luma_weight_op(dest_y, h->mb_linesize, height, h->luma_log2_weight_denom,
 685                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
 686         if(h->use_weight_chroma){
 687             chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
 688                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
 689             chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
 690                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
 691         }
 692     }
 693 }
 694
 695 static av_always_inline void
 696 mc_part(H264Context *h, int n, int square, int height, int delta,
 697         uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 698         int x_offset, int y_offset,
 699         qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 700         qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 701         h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 702         int list0, int list1, int pixel_shift, int chroma_idc)
 703 {
 704     if((h->use_weight==2 && list0 && list1
 705         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
 706        || h->use_weight==1)
 707         mc_part_weighted(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
 708                          x_offset, y_offset, qpix_put, chroma_put,
 709                          weight_op[0], weight_op[1], weight_avg[0],
 710                          weight_avg[1], list0, list1, pixel_shift, chroma_idc);
 711     else
 712         mc_part_std(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
 713                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
 714                     chroma_avg, list0, list1, pixel_shift, chroma_idc);
 715 }
 716
 717 static av_always_inline void
 718 prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma_idc)
 719 {
 720     /* fetch pixels for estimated mv 4 macroblocks ahead
 721      * optimized for 64byte cache lines */
 722     MpegEncContext * const s = &h->s;
 723     const int refn = h->ref_cache[list][scan8[0]];
 724     if(refn >= 0){
 725         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
 726         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
 727         uint8_t **src = h->ref_list[list][refn].f.data;
 728         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
 729         s->dsp.prefetch(src[0]+off, s->linesize, 4);
 730         if (chroma_idc == 3 /* yuv444 */) {
 731             s->dsp.prefetch(src[1]+off, s->linesize, 4);
 732             s->dsp.prefetch(src[2]+off, s->linesize, 4);
 733         }else{
 734             off= (((mx>>1)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
 735             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
 736         }
 737     }
 738 }
 739
 740 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 741                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 742                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 743                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 744                       int pixel_shift, int chroma_idc)
 745 {
 746     MpegEncContext * const s = &h->s;
 747     const int mb_xy= h->mb_xy;
 748     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 749
 750     assert(IS_INTER(mb_type));
 751
 752     if(HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
 753         await_references(h);
 754     prefetch_motion(h, 0, pixel_shift, chroma_idc);
 755
 756     if(IS_16X16(mb_type)){
 757         mc_part(h, 0, 1, 16, 0, dest_y, dest_cb, dest_cr, 0, 0,
 758                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
 759                 weight_op, weight_avg,
 760                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 761                 pixel_shift, chroma_idc);
 762     }else if(IS_16X8(mb_type)){
 763         mc_part(h, 0, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
 764                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 765                 weight_op, weight_avg,
 766                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 767                 pixel_shift, chroma_idc);
 768         mc_part(h, 8, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
 769                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 770                 weight_op, weight_avg,
 771                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 772                 pixel_shift, chroma_idc);
 773     }else if(IS_8X16(mb_type)){
 774         mc_part(h, 0, 0, 16, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
 775                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 776                 &weight_op[1], &weight_avg[1],
 777                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 778                 pixel_shift, chroma_idc);
 779         mc_part(h, 4, 0, 16, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
 780                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 781                 &weight_op[1], &weight_avg[1],
 782                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 783                 pixel_shift, chroma_idc);
 784     }else{
 785         int i;
 786
 787         assert(IS_8X8(mb_type));
 788
 789         for(i=0; i<4; i++){
 790             const int sub_mb_type= h->sub_mb_type[i];
 791             const int n= 4*i;
 792             int x_offset= (i&1)<<2;
 793             int y_offset= (i&2)<<1;
 794
 795             if(IS_SUB_8X8(sub_mb_type)){
 796                 mc_part(h, n, 1, 8, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 797                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 798                     &weight_op[1], &weight_avg[1],
 799                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 800                     pixel_shift, chroma_idc);
 801             }else if(IS_SUB_8X4(sub_mb_type)){
 802                 mc_part(h, n  , 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 803                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 804                     &weight_op[1], &weight_avg[1],
 805                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 806                     pixel_shift, chroma_idc);
 807                 mc_part(h, n+2, 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
 808                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 809                     &weight_op[1], &weight_avg[1],
 810                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 811                     pixel_shift, chroma_idc);
 812             }else if(IS_SUB_4X8(sub_mb_type)){
 813                 mc_part(h, n  , 0, 8, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 814                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 815                     &weight_op[2], &weight_avg[2],
 816                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 817                     pixel_shift, chroma_idc);
 818                 mc_part(h, n+1, 0, 8, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
 819                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 820                     &weight_op[2], &weight_avg[2],
 821                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 822                     pixel_shift, chroma_idc);
 823             }else{
 824                 int j;
 825                 assert(IS_SUB_4X4(sub_mb_type));
 826                 for(j=0; j<4; j++){
 827                     int sub_x_offset= x_offset + 2*(j&1);
 828                     int sub_y_offset= y_offset +   (j&2);
 829                     mc_part(h, n+j, 1, 4, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
 830                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 831                         &weight_op[2], &weight_avg[2],
 832                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 833                         pixel_shift, chroma_idc);
 834                 }
 835             }
 836         }
 837     }
 838
 839     prefetch_motion(h, 1, pixel_shift, chroma_idc);
 840 }
 841
 842 static av_always_inline void
 843 hl_motion_420(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 844               qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 845               qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 846               h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 847               int pixel_shift)
 848 {
 849     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
 850               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 1);
 851 }
 852
 853 static av_always_inline void
 854 hl_motion_422(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 855               qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 856               qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 857               h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 858               int pixel_shift)
 859 {
 860     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
 861               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 2);
 862 }
 863
 864 static void free_tables(H264Context *h, int free_rbsp){
 865     int i;
 866     H264Context *hx;
 867
 868     av_freep(&h->intra4x4_pred_mode);
 869     av_freep(&h->chroma_pred_mode_table);
 870     av_freep(&h->cbp_table);
 871     av_freep(&h->mvd_table[0]);
 872     av_freep(&h->mvd_table[1]);
 873     av_freep(&h->direct_table);
 874     av_freep(&h->non_zero_count);
 875     av_freep(&h->slice_table_base);
 876     h->slice_table= NULL;
 877     av_freep(&h->list_counts);
 878
 879     av_freep(&h->mb2b_xy);
 880     av_freep(&h->mb2br_xy);
 881
 882     for(i = 0; i < MAX_THREADS; i++) {
 883         hx = h->thread_context[i];
 884         if(!hx) continue;
 885         av_freep(&hx->top_borders[1]);
 886         av_freep(&hx->top_borders[0]);
 887         av_freep(&hx->s.obmc_scratchpad);
 888         if (free_rbsp){
 889             av_freep(&hx->rbsp_buffer[1]);
 890             av_freep(&hx->rbsp_buffer[0]);
 891             hx->rbsp_buffer_size[0] = 0;
 892             hx->rbsp_buffer_size[1] = 0;
 893         }
 894         if (i) av_freep(&h->thread_context[i]);
 895     }
 896 }
 897
 898 static void init_dequant8_coeff_table(H264Context *h){
 899     int i,j,q,x;
 900     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 901
 902     for(i=0; i<6; i++ ){
 903         h->dequant8_coeff[i] = h->dequant8_buffer[i];
 904         for(j=0; j<i; j++){
 905             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
 906                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
 907                 break;
 908             }
 909         }
 910         if(j<i)
 911             continue;
 912
 913         for(q=0; q<max_qp+1; q++){
 914             int shift = div6[q];
 915             int idx = rem6[q];
 916             for(x=0; x<64; x++)
 917                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
 918                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
 919                     h->pps.scaling_matrix8[i][x]) << shift;
 920         }
 921     }
 922 }
 923
 924 static void init_dequant4_coeff_table(H264Context *h){
 925     int i,j,q,x;
 926     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 927     for(i=0; i<6; i++ ){
 928         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 929         for(j=0; j<i; j++){
 930             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
 931                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 932                 break;
 933             }
 934         }
 935         if(j<i)
 936             continue;
 937
 938         for(q=0; q<max_qp+1; q++){
 939             int shift = div6[q] + 2;
 940             int idx = rem6[q];
 941             for(x=0; x<16; x++)
 942                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
 943                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
 944                     h->pps.scaling_matrix4[i][x]) << shift;
 945         }
 946     }
 947 }
 948
 949 static void init_dequant_tables(H264Context *h){
 950     int i,x;
 951     init_dequant4_coeff_table(h);
 952     if(h->pps.transform_8x8_mode)
 953         init_dequant8_coeff_table(h);
 954     if(h->sps.transform_bypass){
 955         for(i=0; i<6; i++)
 956             for(x=0; x<16; x++)
 957                 h->dequant4_coeff[i][0][x] = 1<<6;
 958         if(h->pps.transform_8x8_mode)
 959             for(i=0; i<6; i++)
 960                 for(x=0; x<64; x++)
 961                     h->dequant8_coeff[i][0][x] = 1<<6;
 962     }
 963 }
 964
 965
 966 int ff_h264_alloc_tables(H264Context *h){
 967     MpegEncContext * const s = &h->s;
 968     const int big_mb_num= s->mb_stride * (s->mb_height+1);
 969     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
 970     int x,y;
 971
 972     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
 973
 974     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
 975     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
 976     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
 977
 978     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
 979     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
 980     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
 981     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
 982     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
 983
 984     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
 985     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
 986
 987     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
 988     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
 989     for(y=0; y<s->mb_height; y++){
 990         for(x=0; x<s->mb_width; x++){
 991             const int mb_xy= x + y*s->mb_stride;
 992             const int b_xy = 4*x + 4*y*h->b_stride;
 993
 994             h->mb2b_xy [mb_xy]= b_xy;
 995             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
 996         }
 997     }
 998
 999     s->obmc_scratchpad = NULL;
1000
1001     if(!h->dequant4_coeff[0])
1002         init_dequant_tables(h);
1003
1004     return 0;
1005 fail:
1006     free_tables(h, 1);
1007     return -1;
1008 }
1009
1010 /**
1011  * Mimic alloc_tables(), but for every context thread.
1012  */
1013 static void clone_tables(H264Context *dst, H264Context *src, int i){
1014     MpegEncContext * const s = &src->s;
1015     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
1016     dst->non_zero_count           = src->non_zero_count;
1017     dst->slice_table              = src->slice_table;
1018     dst->cbp_table                = src->cbp_table;
1019     dst->mb2b_xy                  = src->mb2b_xy;
1020     dst->mb2br_xy                 = src->mb2br_xy;
1021     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
1022     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
1023     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
1024     dst->direct_table             = src->direct_table;
1025     dst->list_counts              = src->list_counts;
1026
1027     dst->s.obmc_scratchpad = NULL;
1028     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma, src->sps.chroma_format_idc);
1029 }
1030
1031 /**
1032  * Init context
1033  * Allocate buffers which are not shared amongst multiple threads.
1034  */
1035 static int context_init(H264Context *h){
1036     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
1037     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
1038
1039     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
1040     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
1041
1042     return 0;
1043 fail:
1044     return -1; // free_tables will clean up for us
1045 }
1046
1047 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
1048
1049 static av_cold void common_init(H264Context *h){
1050     MpegEncContext * const s = &h->s;
1051
1052     s->width = s->avctx->width;
1053     s->height = s->avctx->height;
1054     s->codec_id= s->avctx->codec->id;
1055
1056     s->avctx->bits_per_raw_sample = 8;
1057     h->cur_chroma_format_idc = 1;
1058
1059     ff_h264dsp_init(&h->h264dsp,
1060                     s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1061     ff_h264_pred_init(&h->hpc, s->codec_id,
1062                       s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1063
1064     h->dequant_coeff_pps= -1;
1065     s->unrestricted_mv=1;
1066
1067     s->dsp.dct_bits = 16;
1068     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
1069
1070     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
1071     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
1072 }
1073
1074 int ff_h264_decode_extradata(H264Context *h, const uint8_t *buf, int size)
1075 {
1076     AVCodecContext *avctx = h->s.avctx;
1077
1078     if(!buf || size <= 0)
1079         return -1;
1080
1081     if(buf[0] == 1){
1082         int i, cnt, nalsize;
1083         const unsigned char *p = buf;
1084
1085         h->is_avc = 1;
1086
1087         if(size < 7) {
1088             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1089             return -1;
1090         }
1091         /* sps and pps in the avcC always have length coded with 2 bytes,
1092            so put a fake nal_length_size = 2 while parsing them */
1093         h->nal_length_size = 2;
1094         // Decode sps from avcC
1095         cnt = *(p+5) & 0x1f; // Number of sps
1096         p += 6;
1097         for (i = 0; i < cnt; i++) {
1098             nalsize = AV_RB16(p) + 2;
1099             if(nalsize > size - (p-buf))
1100                 return -1;
1101             if(decode_nal_units(h, p, nalsize) < 0) {
1102                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1103                 return -1;
1104             }
1105             p += nalsize;
1106         }
1107         // Decode pps from avcC
1108         cnt = *(p++); // Number of pps
1109         for (i = 0; i < cnt; i++) {
1110             nalsize = AV_RB16(p) + 2;
1111             if(nalsize > size - (p-buf))
1112                 return -1;
1113             if (decode_nal_units(h, p, nalsize) < 0) {
1114                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1115                 return -1;
1116             }
1117             p += nalsize;
1118         }
1119         // Now store right nal length size, that will be use to parse all other nals
1120         h->nal_length_size = (buf[4] & 0x03) + 1;
1121     } else {
1122         h->is_avc = 0;
1123         if(decode_nal_units(h, buf, size) < 0)
1124             return -1;
1125     }
1126     return 0;
1127 }
1128
1129 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1130     H264Context *h= avctx->priv_data;
1131     MpegEncContext * const s = &h->s;
1132     int i;
1133
1134     MPV_decode_defaults(s);
1135
1136     s->avctx = avctx;
1137     common_init(h);
1138
1139     s->out_format = FMT_H264;
1140     s->workaround_bugs= avctx->workaround_bugs;
1141
1142     // set defaults
1143 //    s->decode_mb= ff_h263_decode_mb;
1144     s->quarter_sample = 1;
1145     if(!avctx->has_b_frames)
1146     s->low_delay= 1;
1147
1148     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1149
1150     ff_h264_decode_init_vlc();
1151
1152     h->pixel_shift = 0;
1153     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1154
1155     h->thread_context[0] = h;
1156     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1157     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1158         h->last_pocs[i] = INT_MIN;
1159     h->prev_poc_msb= 1<<16;
1160     h->prev_frame_num= -1;
1161     h->x264_build = -1;
1162     ff_h264_reset_sei(h);
1163     if(avctx->codec_id == CODEC_ID_H264){
1164         if(avctx->ticks_per_frame == 1){
1165             s->avctx->time_base.den *=2;
1166         }
1167         avctx->ticks_per_frame = 2;
1168     }
1169
1170     if(avctx->extradata_size > 0 && avctx->extradata &&
1171         ff_h264_decode_extradata(h, avctx->extradata, avctx->extradata_size))
1172         return -1;
1173
1174     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1175         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1176         s->low_delay = 0;
1177     }
1178
1179     return 0;
1180 }
1181
1182 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1183 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1184 {
1185     int i;
1186
1187     for (i=0; i<count; i++){
1188         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1189                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1190                 !from[i]));
1191         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1192     }
1193 }
1194
1195 static void copy_parameter_set(void **to, void **from, int count, int size)
1196 {
1197     int i;
1198
1199     for (i=0; i<count; i++){
1200         if (to[i] && !from[i]) av_freep(&to[i]);
1201         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1202
1203         if (from[i]) memcpy(to[i], from[i], size);
1204     }
1205 }
1206
1207 static int decode_init_thread_copy(AVCodecContext *avctx){
1208     H264Context *h= avctx->priv_data;
1209
1210     if (!avctx->internal->is_copy)
1211         return 0;
1212     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1213     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1214
1215     return 0;
1216 }
1217
1218 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1219 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1220     H264Context *h= dst->priv_data, *h1= src->priv_data;
1221     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1222     int inited = s->context_initialized, err;
1223     int i;
1224
1225     if(dst == src || !s1->context_initialized) return 0;
1226
1227     err = ff_mpeg_update_thread_context(dst, src);
1228     if(err) return err;
1229
1230     //FIXME handle width/height changing
1231     if(!inited){
1232         for(i = 0; i < MAX_SPS_COUNT; i++)
1233             av_freep(h->sps_buffers + i);
1234
1235         for(i = 0; i < MAX_PPS_COUNT; i++)
1236             av_freep(h->pps_buffers + i);
1237
1238         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1239         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1240         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1241         if (ff_h264_alloc_tables(h) < 0) {
1242             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1243             return AVERROR(ENOMEM);
1244         }
1245         context_init(h);
1246
1247         for(i=0; i<2; i++){
1248             h->rbsp_buffer[i] = NULL;
1249             h->rbsp_buffer_size[i] = 0;
1250         }
1251
1252         h->thread_context[0] = h;
1253
1254         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1255         // so this has to be allocated here
1256         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1257
1258         s->dsp.clear_blocks(h->mb);
1259         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1260     }
1261
1262     //extradata/NAL handling
1263     h->is_avc          = h1->is_avc;
1264
1265     //SPS/PPS
1266     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1267     h->sps             = h1->sps;
1268     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1269     h->pps             = h1->pps;
1270
1271     //Dequantization matrices
1272     //FIXME these are big - can they be only copied when PPS changes?
1273     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1274
1275     for(i=0; i<6; i++)
1276         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1277
1278     for(i=0; i<6; i++)
1279         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1280
1281     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1282
1283     //POC timing
1284     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1285
1286     //reference lists
1287     copy_fields(h, h1, ref_count, list_count);
1288     copy_fields(h, h1, ref_list,  intra_gb);
1289     copy_fields(h, h1, short_ref, cabac_init_idc);
1290
1291     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1292     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1293     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1294
1295     h->last_slice_type = h1->last_slice_type;
1296     h->sync            = h1->sync;
1297
1298     if(!s->current_picture_ptr) return 0;
1299
1300     if(!s->dropable) {
1301         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1302         h->prev_poc_msb     = h->poc_msb;
1303         h->prev_poc_lsb     = h->poc_lsb;
1304     }
1305     h->prev_frame_num_offset= h->frame_num_offset;
1306     h->prev_frame_num       = h->frame_num;
1307     h->outputed_poc         = h->next_outputed_poc;
1308
1309     return err;
1310 }
1311
1312 int ff_h264_frame_start(H264Context *h){
1313     MpegEncContext * const s = &h->s;
1314     int i;
1315     const int pixel_shift = h->pixel_shift;
1316     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1317
1318     if(MPV_frame_start(s, s->avctx) < 0)
1319         return -1;
1320     ff_er_frame_start(s);
1321     /*
1322      * MPV_frame_start uses pict_type to derive key_frame.
1323      * This is incorrect for H.264; IDR markings must be used.
1324      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1325      * See decode_nal_units().
1326      */
1327     s->current_picture_ptr->f.key_frame = 0;
1328     s->current_picture_ptr->mmco_reset= 0;
1329
1330     assert(s->linesize && s->uvlinesize);
1331
1332     for(i=0; i<16; i++){
1333         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1334         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1335     }
1336     for(i=0; i<16; i++){
1337         h->block_offset[16+i]=
1338         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1339         h->block_offset[48+16+i]=
1340         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1341     }
1342
1343     /* can't be in alloc_tables because linesize isn't known there.
1344      * FIXME: redo bipred weight to not require extra buffer? */
1345     for(i = 0; i < thread_count; i++)
1346         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1347             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1348
1349     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1350     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1351
1352 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1353
1354     // We mark the current picture as non-reference after allocating it, so
1355     // that if we break out due to an error it can be released automatically
1356     // in the next MPV_frame_start().
1357     // SVQ3 as well as most other codecs have only last/next/current and thus
1358     // get released even with set reference, besides SVQ3 and others do not
1359     // mark frames as reference later "naturally".
1360     if(s->codec_id != CODEC_ID_SVQ3)
1361         s->current_picture_ptr->f.reference = 0;
1362
1363     s->current_picture_ptr->field_poc[0]=
1364     s->current_picture_ptr->field_poc[1]= INT_MAX;
1365
1366     h->next_output_pic = NULL;
1367
1368     assert(s->current_picture_ptr->long_ref==0);
1369
1370     return 0;
1371 }
1372
1373 /**
1374   * Run setup operations that must be run after slice header decoding.
1375   * This includes finding the next displayed frame.
1376   *
1377   * @param h h264 master context
1378   * @param setup_finished enough NALs have been read that we can call
1379   * ff_thread_finish_setup()
1380   */
1381 static void decode_postinit(H264Context *h, int setup_finished){
1382     MpegEncContext * const s = &h->s;
1383     Picture *out = s->current_picture_ptr;
1384     Picture *cur = s->current_picture_ptr;
1385     int i, pics, out_of_order, out_idx;
1386
1387     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1388     s->current_picture_ptr->f.pict_type   = s->pict_type;
1389
1390     if (h->next_output_pic) return;
1391
1392     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1393         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1394         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1395         //to find this yet, so we assume the worst for now.
1396         //if (setup_finished)
1397         //    ff_thread_finish_setup(s->avctx);
1398         return;
1399     }
1400
1401     cur->f.interlaced_frame = 0;
1402     cur->f.repeat_pict      = 0;
1403
1404     /* Signal interlacing information externally. */
1405     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1406
1407     if(h->sps.pic_struct_present_flag){
1408         switch (h->sei_pic_struct)
1409         {
1410         case SEI_PIC_STRUCT_FRAME:
1411             break;
1412         case SEI_PIC_STRUCT_TOP_FIELD:
1413         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1414             cur->f.interlaced_frame = 1;
1415             break;
1416         case SEI_PIC_STRUCT_TOP_BOTTOM:
1417         case SEI_PIC_STRUCT_BOTTOM_TOP:
1418             if (FIELD_OR_MBAFF_PICTURE)
1419                 cur->f.interlaced_frame = 1;
1420             else
1421                 // try to flag soft telecine progressive
1422                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1423             break;
1424         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1425         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1426             // Signal the possibility of telecined film externally (pic_struct 5,6)
1427             // From these hints, let the applications decide if they apply deinterlacing.
1428             cur->f.repeat_pict = 1;
1429             break;
1430         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1431             // Force progressive here, as doubling interlaced frame is a bad idea.
1432             cur->f.repeat_pict = 2;
1433             break;
1434         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1435             cur->f.repeat_pict = 4;
1436             break;
1437         }
1438
1439         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1440             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1441     }else{
1442         /* Derive interlacing flag from used decoding process. */
1443         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1444     }
1445     h->prev_interlaced_frame = cur->f.interlaced_frame;
1446
1447     if (cur->field_poc[0] != cur->field_poc[1]){
1448         /* Derive top_field_first from field pocs. */
1449         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1450     }else{
1451         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1452             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1453             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1454               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1455                 cur->f.top_field_first = 1;
1456             else
1457                 cur->f.top_field_first = 0;
1458         }else{
1459             /* Most likely progressive */
1460             cur->f.top_field_first = 0;
1461         }
1462     }
1463
1464     cur->mmco_reset = h->mmco_reset;
1465     h->mmco_reset = 0;
1466     //FIXME do something with unavailable reference frames
1467
1468     /* Sort B-frames into display order */
1469
1470     if(h->sps.bitstream_restriction_flag
1471        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1472         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1473         s->low_delay = 0;
1474     }
1475
1476     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1477        && !h->sps.bitstream_restriction_flag){
1478         s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1479         s->low_delay= 0;
1480     }
1481
1482     for (i = 0; 1; i++) {
1483         if(i == MAX_DELAYED_PIC_COUNT || cur->poc < h->last_pocs[i]){
1484             if(i)
1485                 h->last_pocs[i-1] = cur->poc;
1486             break;
1487         } else if(i) {
1488             h->last_pocs[i-1]= h->last_pocs[i];
1489         }
1490     }
1491     out_of_order = MAX_DELAYED_PIC_COUNT - i;
1492     if(   cur->f.pict_type == AV_PICTURE_TYPE_B
1493        || (h->last_pocs[MAX_DELAYED_PIC_COUNT-2] > INT_MIN && h->last_pocs[MAX_DELAYED_PIC_COUNT-1] - h->last_pocs[MAX_DELAYED_PIC_COUNT-2] > 2))
1494         out_of_order = FFMAX(out_of_order, 1);
1495     if(s->avctx->has_b_frames < out_of_order && !h->sps.bitstream_restriction_flag){
1496         av_log(s->avctx, AV_LOG_WARNING, "Increasing reorder buffer to %d\n", out_of_order);
1497         s->avctx->has_b_frames = out_of_order;
1498         s->low_delay = 0;
1499     }
1500
1501     pics = 0;
1502     while(h->delayed_pic[pics]) pics++;
1503
1504     av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
1505
1506     h->delayed_pic[pics++] = cur;
1507     if (cur->f.reference == 0)
1508         cur->f.reference = DELAYED_PIC_REF;
1509
1510     out = h->delayed_pic[0];
1511     out_idx = 0;
1512     for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1513         if(h->delayed_pic[i]->poc < out->poc){
1514             out = h->delayed_pic[i];
1515             out_idx = i;
1516         }
1517     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1518         h->next_outputed_poc= INT_MIN;
1519     out_of_order = out->poc < h->next_outputed_poc;
1520
1521     if(out_of_order || pics > s->avctx->has_b_frames){
1522         out->f.reference &= ~DELAYED_PIC_REF;
1523         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1524                          // or else the first thread can release the picture and reuse it unsafely
1525         for(i=out_idx; h->delayed_pic[i]; i++)
1526             h->delayed_pic[i] = h->delayed_pic[i+1];
1527     }
1528     if(!out_of_order && pics > s->avctx->has_b_frames){
1529         h->next_output_pic = out;
1530         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1531             h->next_outputed_poc = INT_MIN;
1532         } else
1533             h->next_outputed_poc = out->poc;
1534     }else{
1535         av_log(s->avctx, AV_LOG_DEBUG, "no picture %s\n", out_of_order ? "ooo" : "");
1536     }
1537
1538     if (h->next_output_pic && h->next_output_pic->sync) {
1539         h->sync |= 2;
1540     }
1541
1542     if (setup_finished)
1543         ff_thread_finish_setup(s->avctx);
1544 }
1545
1546 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1547                                               uint8_t *src_cb, uint8_t *src_cr,
1548                                               int linesize, int uvlinesize, int simple)
1549 {
1550     MpegEncContext * const s = &h->s;
1551     uint8_t *top_border;
1552     int top_idx = 1;
1553     const int pixel_shift = h->pixel_shift;
1554     int chroma444 = CHROMA444;
1555     int chroma422 = CHROMA422;
1556
1557     src_y  -=   linesize;
1558     src_cb -= uvlinesize;
1559     src_cr -= uvlinesize;
1560
1561     if(!simple && FRAME_MBAFF){
1562         if(s->mb_y&1){
1563             if(!MB_MBAFF){
1564                 top_border = h->top_borders[0][s->mb_x];
1565                 AV_COPY128(top_border, src_y + 15*linesize);
1566                 if (pixel_shift)
1567                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1568                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1569                     if(chroma444){
1570                         if (pixel_shift){
1571                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1572                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1573                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1574                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1575                         } else {
1576                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1577                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1578                         }
1579                     } else if(chroma422){
1580                         if (pixel_shift) {
1581                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1582                             AV_COPY128(top_border+48, src_cr + 15*uvlinesize);
1583                         } else {
1584                             AV_COPY64(top_border+16, src_cb +  15*uvlinesize);
1585                             AV_COPY64(top_border+24, src_cr +  15*uvlinesize);
1586                         }
1587                     } else {
1588                         if (pixel_shift) {
1589                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1590                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1591                         } else {
1592                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1593                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1594                         }
1595                     }
1596                 }
1597             }
1598         }else if(MB_MBAFF){
1599             top_idx = 0;
1600         }else
1601             return;
1602     }
1603
1604     top_border = h->top_borders[top_idx][s->mb_x];
1605     // There are two lines saved, the line above the the top macroblock of a pair,
1606     // and the line above the bottom macroblock
1607     AV_COPY128(top_border, src_y + 16*linesize);
1608     if (pixel_shift)
1609         AV_COPY128(top_border+16, src_y+16*linesize+16);
1610
1611     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1612         if(chroma444){
1613             if (pixel_shift){
1614                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1615                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1616                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1617                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1618             } else {
1619                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1620                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1621             }
1622         } else if(chroma422) {
1623             if (pixel_shift) {
1624                 AV_COPY128(top_border+32, src_cb+16*uvlinesize);
1625                 AV_COPY128(top_border+48, src_cr+16*uvlinesize);
1626             } else {
1627                 AV_COPY64(top_border+16, src_cb+16*uvlinesize);
1628                 AV_COPY64(top_border+24, src_cr+16*uvlinesize);
1629             }
1630         } else {
1631             if (pixel_shift) {
1632                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1633                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1634             } else {
1635                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1636                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1637             }
1638         }
1639     }
1640 }
1641
1642 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1643                                   uint8_t *src_cb, uint8_t *src_cr,
1644                                   int linesize, int uvlinesize,
1645                                   int xchg, int chroma444,
1646                                   int simple, int pixel_shift){
1647     MpegEncContext * const s = &h->s;
1648     int deblock_topleft;
1649     int deblock_top;
1650     int top_idx = 1;
1651     uint8_t *top_border_m1;
1652     uint8_t *top_border;
1653
1654     if(!simple && FRAME_MBAFF){
1655         if(s->mb_y&1){
1656             if(!MB_MBAFF)
1657                 return;
1658         }else{
1659             top_idx = MB_MBAFF ? 0 : 1;
1660         }
1661     }
1662
1663     if(h->deblocking_filter == 2) {
1664         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1665         deblock_top     = h->top_type;
1666     } else {
1667         deblock_topleft = (s->mb_x > 0);
1668         deblock_top     = (s->mb_y > !!MB_FIELD);
1669     }
1670
1671     src_y  -=   linesize + 1 + pixel_shift;
1672     src_cb -= uvlinesize + 1 + pixel_shift;
1673     src_cr -= uvlinesize + 1 + pixel_shift;
1674
1675     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1676     top_border    = h->top_borders[top_idx][s->mb_x];
1677
1678 #define XCHG(a,b,xchg)\
1679     if (pixel_shift) {\
1680         if (xchg) {\
1681             AV_SWAP64(b+0,a+0);\
1682             AV_SWAP64(b+8,a+8);\
1683         } else {\
1684             AV_COPY128(b,a); \
1685         }\
1686     } else \
1687 if (xchg) AV_SWAP64(b,a);\
1688 else      AV_COPY64(b,a);
1689
1690     if(deblock_top){
1691         if(deblock_topleft){
1692             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1693         }
1694         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1695         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1696         if(s->mb_x+1 < s->mb_width){
1697             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1698         }
1699     }
1700     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1701         if(chroma444){
1702             if(deblock_topleft){
1703                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1704                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1705             }
1706             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1707             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1708             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1709             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1710             if(s->mb_x+1 < s->mb_width){
1711                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1712                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1713             }
1714         } else {
1715             if(deblock_top){
1716                 if(deblock_topleft){
1717                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1718                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1719                 }
1720                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1721                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1722             }
1723         }
1724     }
1725 }
1726
1727 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1728     if (high_bit_depth) {
1729         return AV_RN32A(((int32_t*)mb) + index);
1730     } else
1731         return AV_RN16A(mb + index);
1732 }
1733
1734 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1735     if (high_bit_depth) {
1736         AV_WN32A(((int32_t*)mb) + index, value);
1737     } else
1738         AV_WN16A(mb + index, value);
1739 }
1740
1741 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1742                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1743 {
1744     MpegEncContext * const s = &h->s;
1745     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1746     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1747     int i;
1748     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1749     block_offset += 16*p;
1750     if(IS_INTRA4x4(mb_type)){
1751         if(simple || !s->encoding){
1752             if(IS_8x8DCT(mb_type)){
1753                 if(transform_bypass){
1754                     idct_dc_add =
1755                     idct_add    = s->dsp.add_pixels8;
1756                 }else{
1757                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1758                     idct_add    = h->h264dsp.h264_idct8_add;
1759                 }
1760                 for(i=0; i<16; i+=4){
1761                     uint8_t * const ptr= dest_y + block_offset[i];
1762                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1763                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1764                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1765                     }else{
1766                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1767                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1768                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1769                         if(nnz){
1770                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1771                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1772                             else
1773                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1774                         }
1775                     }
1776                 }
1777             }else{
1778                 if(transform_bypass){
1779                     idct_dc_add =
1780                     idct_add    = s->dsp.add_pixels4;
1781                 }else{
1782                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1783                     idct_add    = h->h264dsp.h264_idct_add;
1784                 }
1785                 for(i=0; i<16; i++){
1786                     uint8_t * const ptr= dest_y + block_offset[i];
1787                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1788
1789                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1790                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1791                     }else{
1792                         uint8_t *topright;
1793                         int nnz, tr;
1794                         uint64_t tr_high;
1795                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1796                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1797                             assert(s->mb_y || linesize <= block_offset[i]);
1798                             if(!topright_avail){
1799                                 if (pixel_shift) {
1800                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1801                                     topright= (uint8_t*) &tr_high;
1802                                 } else {
1803                                     tr= ptr[3 - linesize]*0x01010101u;
1804                                     topright= (uint8_t*) &tr;
1805                                 }
1806                             }else
1807                                 topright= ptr + (4 << pixel_shift) - linesize;
1808                         }else
1809                             topright= NULL;
1810
1811                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1812                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1813                         if(nnz){
1814                             if(is_h264){
1815                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1816                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1817                                 else
1818                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1819                             }else
1820                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1821                         }
1822                     }
1823                 }
1824             }
1825         }
1826     }else{
1827         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1828         if(is_h264){
1829             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1830                 if(!transform_bypass)
1831                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1832                 else{
1833                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1834                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1835                     for(i = 0; i < 16; i++)
1836                         dctcoef_set(h->mb+(p*256 << pixel_shift), pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1837                 }
1838             }
1839         }else
1840             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1841     }
1842 }
1843
1844 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1845                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1846 {
1847     MpegEncContext * const s = &h->s;
1848     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1849     int i;
1850     block_offset += 16*p;
1851     if(!IS_INTRA4x4(mb_type)){
1852         if(is_h264){
1853             if(IS_INTRA16x16(mb_type)){
1854                 if(transform_bypass){
1855                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1856                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1857                     }else{
1858                         for(i=0; i<16; i++){
1859                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1860                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1861                         }
1862                     }
1863                 }else{
1864                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1865                 }
1866             }else if(h->cbp&15){
1867                 if(transform_bypass){
1868                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1869                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1870                     for(i=0; i<16; i+=di){
1871                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1872                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1873                         }
1874                     }
1875                 }else{
1876                     if(IS_8x8DCT(mb_type)){
1877                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1878                     }else{
1879                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1880                     }
1881                 }
1882             }
1883         }else{
1884             for(i=0; i<16; i++){
1885                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1886                     uint8_t * const ptr= dest_y + block_offset[i];
1887                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1888                 }
1889             }
1890         }
1891     }
1892 }
1893
1894 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift)
1895 {
1896     MpegEncContext * const s = &h->s;
1897     const int mb_x= s->mb_x;
1898     const int mb_y= s->mb_y;
1899     const int mb_xy= h->mb_xy;
1900     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1901     uint8_t  *dest_y, *dest_cb, *dest_cr;
1902     int linesize, uvlinesize /*dct_offset*/;
1903     int i, j;
1904     int *block_offset = &h->block_offset[0];
1905     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1906     /* is_h264 should always be true if SVQ3 is disabled. */
1907     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1908     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1909     const int block_h = 16 >> s->chroma_y_shift;
1910     const int chroma422 = CHROMA422;
1911
1912     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1913     dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1914     dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1915
1916     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1917     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1918
1919     h->list_counts[mb_xy]= h->list_count;
1920
1921     if (!simple && MB_FIELD) {
1922         linesize   = h->mb_linesize   = s->linesize * 2;
1923         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1924         block_offset = &h->block_offset[48];
1925         if(mb_y&1){ //FIXME move out of this function?
1926             dest_y -= s->linesize*15;
1927             dest_cb-= s->uvlinesize * (block_h - 1);
1928             dest_cr-= s->uvlinesize * (block_h - 1);
1929         }
1930         if(FRAME_MBAFF) {
1931             int list;
1932             for(list=0; list<h->list_count; list++){
1933                 if(!USES_LIST(mb_type, list))
1934                     continue;
1935                 if(IS_16X16(mb_type)){
1936                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1937                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1938                 }else{
1939                     for(i=0; i<16; i+=4){
1940                         int ref = h->ref_cache[list][scan8[i]];
1941                         if(ref >= 0)
1942                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1943                     }
1944                 }
1945             }
1946         }
1947     } else {
1948         linesize   = h->mb_linesize   = s->linesize;
1949         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1950 //        dct_offset = s->linesize * 16;
1951     }
1952
1953     if (!simple && IS_INTRA_PCM(mb_type)) {
1954         const int bit_depth = h->sps.bit_depth_luma;
1955         if (pixel_shift) {
1956             int j;
1957             GetBitContext gb;
1958             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1959
1960             for (i = 0; i < 16; i++) {
1961                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1962                 for (j = 0; j < 16; j++)
1963                     tmp_y[j] = get_bits(&gb, bit_depth);
1964             }
1965             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1966                 if (!h->sps.chroma_format_idc) {
1967                     for (i = 0; i < block_h; i++) {
1968                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1969                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1970                         for (j = 0; j < 8; j++) {
1971                             tmp_cb[j] = tmp_cr[j] = 1 << (bit_depth - 1);
1972                         }
1973                     }
1974                 } else {
1975                     for (i = 0; i < block_h; i++) {
1976                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1977                         for (j = 0; j < 8; j++)
1978                             tmp_cb[j] = get_bits(&gb, bit_depth);
1979                     }
1980                     for (i = 0; i < block_h; i++) {
1981                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1982                         for (j = 0; j < 8; j++)
1983                             tmp_cr[j] = get_bits(&gb, bit_depth);
1984                     }
1985                 }
1986             }
1987         } else {
1988             for (i=0; i<16; i++) {
1989                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1990             }
1991             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1992                 if (!h->sps.chroma_format_idc) {
1993                     for (i=0; i<8; i++) {
1994                         memset(dest_cb + i*uvlinesize, 1 << (bit_depth - 1), 8);
1995                         memset(dest_cr + i*uvlinesize, 1 << (bit_depth - 1), 8);
1996                     }
1997                 } else {
1998                     for (i=0; i<block_h; i++) {
1999                         memcpy(dest_cb + i*uvlinesize, h->mb + 128 + i*4,  8);
2000                         memcpy(dest_cr + i*uvlinesize, h->mb + 160 + i*4,  8);
2001                     }
2002                 }
2003             }
2004         }
2005     } else {
2006         if(IS_INTRA(mb_type)){
2007             if(h->deblocking_filter)
2008                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
2009
2010             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2011                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2012                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2013             }
2014
2015             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
2016
2017             if(h->deblocking_filter)
2018                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
2019         }else if(is_h264){
2020             if (chroma422) {
2021                 hl_motion_422(h, dest_y, dest_cb, dest_cr,
2022                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2023                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2024                               h->h264dsp.weight_h264_pixels_tab,
2025                               h->h264dsp.biweight_h264_pixels_tab,
2026                               pixel_shift);
2027             } else {
2028                 hl_motion_420(h, dest_y, dest_cb, dest_cr,
2029                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2030                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2031                               h->h264dsp.weight_h264_pixels_tab,
2032                               h->h264dsp.biweight_h264_pixels_tab,
2033                               pixel_shift);
2034             }
2035         }
2036
2037         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
2038
2039         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
2040             uint8_t *dest[2] = {dest_cb, dest_cr};
2041             if(transform_bypass){
2042                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
2043                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
2044                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
2045                 }else{
2046                     idct_add = s->dsp.add_pixels4;
2047                     for(j=1; j<3; j++){
2048                         for(i=j*16; i<j*16+4; i++){
2049                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
2050                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
2051                         }
2052                         if (chroma422) {
2053                             for(i=j*16+4; i<j*16+8; i++){
2054                                 if(h->non_zero_count_cache[ scan8[i+4] ] || dctcoef_get(h->mb, pixel_shift, i*16))
2055                                     idct_add   (dest[j-1] + block_offset[i+4], h->mb + (i*16 << pixel_shift), uvlinesize);
2056                             }
2057                         }
2058                     }
2059                 }
2060             }else{
2061                 if(is_h264){
2062                     int qp[2];
2063                     if (chroma422) {
2064                         qp[0] = h->chroma_qp[0] + 3;
2065                         qp[1] = h->chroma_qp[1] + 3;
2066                     } else {
2067                         qp[0] = h->chroma_qp[0];
2068                         qp[1] = h->chroma_qp[1];
2069                     }
2070                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
2071                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*1 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][qp[0]][0]);
2072                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
2073                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*2 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][qp[1]][0]);
2074                     h->h264dsp.h264_idct_add8(dest, block_offset,
2075                                               h->mb, uvlinesize,
2076                                               h->non_zero_count_cache);
2077                 }
2078 #if CONFIG_SVQ3_DECODER
2079                 else{
2080                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16*16*1, h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
2081                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16*16*2, h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
2082                     for(j=1; j<3; j++){
2083                         for(i=j*16; i<j*16+4; i++){
2084                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2085                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
2086                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
2087                             }
2088                         }
2089                     }
2090                 }
2091 #endif
2092             }
2093         }
2094     }
2095     if(h->cbp || IS_INTRA(mb_type))
2096     {
2097         s->dsp.clear_blocks(h->mb);
2098         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2099     }
2100 }
2101
2102 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
2103     MpegEncContext * const s = &h->s;
2104     const int mb_x= s->mb_x;
2105     const int mb_y= s->mb_y;
2106     const int mb_xy= h->mb_xy;
2107     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2108     uint8_t  *dest[3];
2109     int linesize;
2110     int i, j, p;
2111     int *block_offset = &h->block_offset[0];
2112     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2113     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
2114
2115     for (p = 0; p < plane_count; p++)
2116     {
2117         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2118         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
2119     }
2120
2121     h->list_counts[mb_xy]= h->list_count;
2122
2123     if (!simple && MB_FIELD) {
2124         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2125         block_offset = &h->block_offset[48];
2126         if(mb_y&1) //FIXME move out of this function?
2127             for (p = 0; p < 3; p++)
2128                 dest[p] -= s->linesize*15;
2129         if(FRAME_MBAFF) {
2130             int list;
2131             for(list=0; list<h->list_count; list++){
2132                 if(!USES_LIST(mb_type, list))
2133                     continue;
2134                 if(IS_16X16(mb_type)){
2135                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2136                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2137                 }else{
2138                     for(i=0; i<16; i+=4){
2139                         int ref = h->ref_cache[list][scan8[i]];
2140                         if(ref >= 0)
2141                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2142                     }
2143                 }
2144             }
2145         }
2146     } else {
2147         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2148     }
2149
2150     if (!simple && IS_INTRA_PCM(mb_type)) {
2151         if (pixel_shift) {
2152             const int bit_depth = h->sps.bit_depth_luma;
2153             GetBitContext gb;
2154             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
2155
2156             for (p = 0; p < plane_count; p++) {
2157                 for (i = 0; i < 16; i++) {
2158                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2159                     for (j = 0; j < 16; j++)
2160                         tmp[j] = get_bits(&gb, bit_depth);
2161                 }
2162             }
2163         } else {
2164             for (p = 0; p < plane_count; p++) {
2165                 for (i = 0; i < 16; i++) {
2166                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2167                 }
2168             }
2169         }
2170     } else {
2171         if(IS_INTRA(mb_type)){
2172             if(h->deblocking_filter)
2173                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2174
2175             for (p = 0; p < plane_count; p++)
2176                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2177
2178             if(h->deblocking_filter)
2179                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2180         }else{
2181             hl_motion(h, dest[0], dest[1], dest[2],
2182                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2183                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2184                       h->h264dsp.weight_h264_pixels_tab,
2185                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 3);
2186         }
2187
2188         for (p = 0; p < plane_count; p++)
2189             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2190     }
2191     if(h->cbp || IS_INTRA(mb_type))
2192     {
2193         s->dsp.clear_blocks(h->mb);
2194         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2195     }
2196 }
2197
2198 /**
2199  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2200  */
2201 #define hl_decode_mb_simple(sh, bits) \
2202 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2203     hl_decode_mb_internal(h, 1, sh); \
2204 }
2205 hl_decode_mb_simple(0, 8)
2206 hl_decode_mb_simple(1, 16)
2207
2208 /**
2209  * Process a macroblock; this handles edge cases, such as interlacing.
2210  */
2211 static void av_noinline hl_decode_mb_complex(H264Context *h){
2212     hl_decode_mb_internal(h, 0, h->pixel_shift);
2213 }
2214
2215 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2216     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2217 }
2218
2219 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2220     hl_decode_mb_444_internal(h, 1, 0);
2221 }
2222
2223 void ff_h264_hl_decode_mb(H264Context *h){
2224     MpegEncContext * const s = &h->s;
2225     const int mb_xy= h->mb_xy;
2226     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2227     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2228
2229     if (CHROMA444) {
2230         if(is_complex || h->pixel_shift)
2231             hl_decode_mb_444_complex(h);
2232         else
2233             hl_decode_mb_444_simple(h);
2234     } else if (is_complex) {
2235         hl_decode_mb_complex(h);
2236     } else if (h->pixel_shift) {
2237         hl_decode_mb_simple_16(h);
2238     } else
2239         hl_decode_mb_simple_8(h);
2240 }
2241
2242 static int pred_weight_table(H264Context *h){
2243     MpegEncContext * const s = &h->s;
2244     int list, i;
2245     int luma_def, chroma_def;
2246
2247     h->use_weight= 0;
2248     h->use_weight_chroma= 0;
2249     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2250     if(h->sps.chroma_format_idc)
2251         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2252     luma_def = 1<<h->luma_log2_weight_denom;
2253     chroma_def = 1<<h->chroma_log2_weight_denom;
2254
2255     for(list=0; list<2; list++){
2256         h->luma_weight_flag[list]   = 0;
2257         h->chroma_weight_flag[list] = 0;
2258         for(i=0; i<h->ref_count[list]; i++){
2259             int luma_weight_flag, chroma_weight_flag;
2260
2261             luma_weight_flag= get_bits1(&s->gb);
2262             if(luma_weight_flag){
2263                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2264                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2265                 if(   h->luma_weight[i][list][0] != luma_def
2266                    || h->luma_weight[i][list][1] != 0) {
2267                     h->use_weight= 1;
2268                     h->luma_weight_flag[list]= 1;
2269                 }
2270             }else{
2271                 h->luma_weight[i][list][0]= luma_def;
2272                 h->luma_weight[i][list][1]= 0;
2273             }
2274
2275             if(h->sps.chroma_format_idc){
2276                 chroma_weight_flag= get_bits1(&s->gb);
2277                 if(chroma_weight_flag){
2278                     int j;
2279                     for(j=0; j<2; j++){
2280                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2281                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2282                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2283                            || h->chroma_weight[i][list][j][1] != 0) {
2284                             h->use_weight_chroma= 1;
2285                             h->chroma_weight_flag[list]= 1;
2286                         }
2287                     }
2288                 }else{
2289                     int j;
2290                     for(j=0; j<2; j++){
2291                         h->chroma_weight[i][list][j][0]= chroma_def;
2292                         h->chroma_weight[i][list][j][1]= 0;
2293                     }
2294                 }
2295             }
2296         }
2297         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2298     }
2299     h->use_weight= h->use_weight || h->use_weight_chroma;
2300     return 0;
2301 }
2302
2303 /**
2304  * Initialize implicit_weight table.
2305  * @param field  0/1 initialize the weight for interlaced MBAFF
2306  *                -1 initializes the rest
2307  */
2308 static void implicit_weight_table(H264Context *h, int field){
2309     MpegEncContext * const s = &h->s;
2310     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2311
2312     for (i = 0; i < 2; i++) {
2313         h->luma_weight_flag[i]   = 0;
2314         h->chroma_weight_flag[i] = 0;
2315     }
2316
2317     if(field < 0){
2318         if (s->picture_structure == PICT_FRAME) {
2319             cur_poc = s->current_picture_ptr->poc;
2320         } else {
2321             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2322         }
2323     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2324        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2325         h->use_weight= 0;
2326         h->use_weight_chroma= 0;
2327         return;
2328     }
2329         ref_start= 0;
2330         ref_count0= h->ref_count[0];
2331         ref_count1= h->ref_count[1];
2332     }else{
2333         cur_poc = s->current_picture_ptr->field_poc[field];
2334         ref_start= 16;
2335         ref_count0= 16+2*h->ref_count[0];
2336         ref_count1= 16+2*h->ref_count[1];
2337     }
2338
2339     h->use_weight= 2;
2340     h->use_weight_chroma= 2;
2341     h->luma_log2_weight_denom= 5;
2342     h->chroma_log2_weight_denom= 5;
2343
2344     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2345         int poc0 = h->ref_list[0][ref0].poc;
2346         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2347             int w = 32;
2348             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2349                 int poc1 = h->ref_list[1][ref1].poc;
2350                 int td = av_clip(poc1 - poc0, -128, 127);
2351                 if(td){
2352                     int tb = av_clip(cur_poc - poc0, -128, 127);
2353                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2354                     int dist_scale_factor = (tb*tx + 32) >> 8;
2355                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2356                         w = 64 - dist_scale_factor;
2357                 }
2358             }
2359             if(field<0){
2360                 h->implicit_weight[ref0][ref1][0]=
2361                 h->implicit_weight[ref0][ref1][1]= w;
2362             }else{
2363                 h->implicit_weight[ref0][ref1][field]=w;
2364             }
2365         }
2366     }
2367 }
2368
2369 /**
2370  * instantaneous decoder refresh.
2371  */
2372 static void idr(H264Context *h){
2373     int i;
2374     ff_h264_remove_all_refs(h);
2375     h->prev_frame_num= 0;
2376     h->prev_frame_num_offset= 0;
2377     h->prev_poc_msb= 1<<16;
2378     h->prev_poc_lsb= 0;
2379     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2380         h->last_pocs[i] = INT_MIN;
2381 }
2382
2383 /* forget old pics after a seek */
2384 static void flush_dpb(AVCodecContext *avctx){
2385     H264Context *h= avctx->priv_data;
2386     int i;
2387     for(i=0; i<=MAX_DELAYED_PIC_COUNT; i++) {
2388         if(h->delayed_pic[i])
2389             h->delayed_pic[i]->f.reference = 0;
2390         h->delayed_pic[i]= NULL;
2391     }
2392     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2393     h->prev_interlaced_frame = 1;
2394     idr(h);
2395     h->prev_frame_num= -1;
2396     if(h->s.current_picture_ptr)
2397         h->s.current_picture_ptr->f.reference = 0;
2398     h->s.first_field= 0;
2399     ff_h264_reset_sei(h);
2400     ff_mpeg_flush(avctx);
2401     h->recovery_frame= -1;
2402     h->sync= 0;
2403 }
2404
2405 static int init_poc(H264Context *h){
2406     MpegEncContext * const s = &h->s;
2407     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2408     int field_poc[2];
2409     Picture *cur = s->current_picture_ptr;
2410
2411     h->frame_num_offset= h->prev_frame_num_offset;
2412     if(h->frame_num < h->prev_frame_num)
2413         h->frame_num_offset += max_frame_num;
2414
2415     if(h->sps.poc_type==0){
2416         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2417
2418         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2419             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2420         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2421             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2422         else
2423             h->poc_msb = h->prev_poc_msb;
2424 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2425         field_poc[0] =
2426         field_poc[1] = h->poc_msb + h->poc_lsb;
2427         if(s->picture_structure == PICT_FRAME)
2428             field_poc[1] += h->delta_poc_bottom;
2429     }else if(h->sps.poc_type==1){
2430         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2431         int i;
2432
2433         if(h->sps.poc_cycle_length != 0)
2434             abs_frame_num = h->frame_num_offset + h->frame_num;
2435         else
2436             abs_frame_num = 0;
2437
2438         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2439             abs_frame_num--;
2440
2441         expected_delta_per_poc_cycle = 0;
2442         for(i=0; i < h->sps.poc_cycle_length; i++)
2443             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2444
2445         if(abs_frame_num > 0){
2446             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2447             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2448
2449             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2450             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2451                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2452         } else
2453             expectedpoc = 0;
2454
2455         if(h->nal_ref_idc == 0)
2456             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2457
2458         field_poc[0] = expectedpoc + h->delta_poc[0];
2459         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2460
2461         if(s->picture_structure == PICT_FRAME)
2462             field_poc[1] += h->delta_poc[1];
2463     }else{
2464         int poc= 2*(h->frame_num_offset + h->frame_num);
2465
2466         if(!h->nal_ref_idc)
2467             poc--;
2468
2469         field_poc[0]= poc;
2470         field_poc[1]= poc;
2471     }
2472
2473     if(s->picture_structure != PICT_BOTTOM_FIELD)
2474         s->current_picture_ptr->field_poc[0]= field_poc[0];
2475     if(s->picture_structure != PICT_TOP_FIELD)
2476         s->current_picture_ptr->field_poc[1]= field_poc[1];
2477     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2478
2479     return 0;
2480 }
2481
2482
2483 /**
2484  * initialize scan tables
2485  */
2486 static void init_scan_tables(H264Context *h){
2487     int i;
2488     for(i=0; i<16; i++){
2489 #define T(x) (x>>2) | ((x<<2) & 0xF)
2490         h->zigzag_scan[i] = T(zigzag_scan[i]);
2491         h-> field_scan[i] = T( field_scan[i]);
2492 #undef T
2493     }
2494     for(i=0; i<64; i++){
2495 #define T(x) (x>>3) | ((x&7)<<3)
2496         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2497         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2498         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2499         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2500 #undef T
2501     }
2502     if(h->sps.transform_bypass){ //FIXME same ugly
2503         h->zigzag_scan_q0          = zigzag_scan;
2504         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2505         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2506         h->field_scan_q0           = field_scan;
2507         h->field_scan8x8_q0        = field_scan8x8;
2508         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2509     }else{
2510         h->zigzag_scan_q0          = h->zigzag_scan;
2511         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2512         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2513         h->field_scan_q0           = h->field_scan;
2514         h->field_scan8x8_q0        = h->field_scan8x8;
2515         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2516     }
2517 }
2518
2519 static int field_end(H264Context *h, int in_setup){
2520     MpegEncContext * const s = &h->s;
2521     AVCodecContext * const avctx= s->avctx;
2522     int err = 0;
2523     s->mb_y= 0;
2524
2525     if (!in_setup && !s->dropable)
2526         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2527                                  s->picture_structure==PICT_BOTTOM_FIELD);
2528
2529     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2530         ff_vdpau_h264_set_reference_frames(s);
2531
2532     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2533         if(!s->dropable) {
2534             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2535             h->prev_poc_msb= h->poc_msb;
2536             h->prev_poc_lsb= h->poc_lsb;
2537         }
2538         h->prev_frame_num_offset= h->frame_num_offset;
2539         h->prev_frame_num= h->frame_num;
2540         h->outputed_poc = h->next_outputed_poc;
2541     }
2542
2543     if (avctx->hwaccel) {
2544         if (avctx->hwaccel->end_frame(avctx) < 0)
2545             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2546     }
2547
2548     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2549         ff_vdpau_h264_picture_complete(s);
2550
2551     /*
2552      * FIXME: Error handling code does not seem to support interlaced
2553      * when slices span multiple rows
2554      * The ff_er_add_slice calls don't work right for bottom
2555      * fields; they cause massive erroneous error concealing
2556      * Error marking covers both fields (top and bottom).
2557      * This causes a mismatched s->error_count
2558      * and a bad error table. Further, the error count goes to
2559      * INT_MAX when called for bottom field, because mb_y is
2560      * past end by one (callers fault) and resync_mb_y != 0
2561      * causes problems for the first MB line, too.
2562      */
2563     if (!FIELD_PICTURE)
2564         ff_er_frame_end(s);
2565
2566     MPV_frame_end(s);
2567
2568     h->current_slice=0;
2569
2570     return err;
2571 }
2572
2573 /**
2574  * Replicate H264 "master" context to thread contexts.
2575  */
2576 static void clone_slice(H264Context *dst, H264Context *src)
2577 {
2578     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2579     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2580     dst->s.current_picture      = src->s.current_picture;
2581     dst->s.linesize             = src->s.linesize;
2582     dst->s.uvlinesize           = src->s.uvlinesize;
2583     dst->s.first_field          = src->s.first_field;
2584
2585     dst->prev_poc_msb           = src->prev_poc_msb;
2586     dst->prev_poc_lsb           = src->prev_poc_lsb;
2587     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2588     dst->prev_frame_num         = src->prev_frame_num;
2589     dst->short_ref_count        = src->short_ref_count;
2590
2591     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2592     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2593     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2594     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2595
2596     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2597     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2598 }
2599
2600 /**
2601  * Compute profile from profile_idc and constraint_set?_flags.
2602  *
2603  * @param sps SPS
2604  *
2605  * @return profile as defined by FF_PROFILE_H264_*
2606  */
2607 int ff_h264_get_profile(SPS *sps)
2608 {
2609     int profile = sps->profile_idc;
2610
2611     switch(sps->profile_idc) {
2612     case FF_PROFILE_H264_BASELINE:
2613         // constraint_set1_flag set to 1
2614         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2615         break;
2616     case FF_PROFILE_H264_HIGH_10:
2617     case FF_PROFILE_H264_HIGH_422:
2618     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2619         // constraint_set3_flag set to 1
2620         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2621         break;
2622     }
2623
2624     return profile;
2625 }
2626
2627 /**
2628  * Decode a slice header.
2629  * This will also call MPV_common_init() and frame_start() as needed.
2630  *
2631  * @param h h264context
2632  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2633  *
2634  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2635  */
2636 static int decode_slice_header(H264Context *h, H264Context *h0){
2637     MpegEncContext * const s = &h->s;
2638     MpegEncContext * const s0 = &h0->s;
2639     unsigned int first_mb_in_slice;
2640     unsigned int pps_id;
2641     int num_ref_idx_active_override_flag;
2642     unsigned int slice_type, tmp, i, j;
2643     int default_ref_list_done = 0;
2644     int last_pic_structure;
2645
2646     s->dropable= h->nal_ref_idc == 0;
2647
2648     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2649     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2650         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2651         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2652     }else{
2653         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2654         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2655     }
2656
2657     first_mb_in_slice= get_ue_golomb_long(&s->gb);
2658
2659     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2660         if(h0->current_slice && FIELD_PICTURE){
2661             field_end(h, 1);
2662         }
2663
2664         h0->current_slice = 0;
2665         if (!s0->first_field)
2666             s->current_picture_ptr= NULL;
2667     }
2668
2669     slice_type= get_ue_golomb_31(&s->gb);
2670     if(slice_type > 9){
2671         av_log(h->s.avctx, AV_LOG_ERROR, "slice type too large (%d) at %d %d\n", h->slice_type, s->mb_x, s->mb_y);
2672         return -1;
2673     }
2674     if(slice_type > 4){
2675         slice_type -= 5;
2676         h->slice_type_fixed=1;
2677     }else
2678         h->slice_type_fixed=0;
2679
2680     slice_type= golomb_to_pict_type[ slice_type ];
2681     if (slice_type == AV_PICTURE_TYPE_I
2682         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2683         default_ref_list_done = 1;
2684     }
2685     h->slice_type= slice_type;
2686     h->slice_type_nos= slice_type & 3;
2687
2688     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2689
2690     pps_id= get_ue_golomb(&s->gb);
2691     if(pps_id>=MAX_PPS_COUNT){
2692         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2693         return -1;
2694     }
2695     if(!h0->pps_buffers[pps_id]) {
2696         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2697         return -1;
2698     }
2699     h->pps= *h0->pps_buffers[pps_id];
2700
2701     if(!h0->sps_buffers[h->pps.sps_id]) {
2702         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2703         return -1;
2704     }
2705     h->sps = *h0->sps_buffers[h->pps.sps_id];
2706
2707     s->avctx->profile = ff_h264_get_profile(&h->sps);
2708     s->avctx->level   = h->sps.level_idc;
2709     s->avctx->refs    = h->sps.ref_frame_count;
2710
2711     if(h == h0 && h->dequant_coeff_pps != pps_id){
2712         h->dequant_coeff_pps = pps_id;
2713         init_dequant_tables(h);
2714     }
2715
2716     s->mb_width= h->sps.mb_width;
2717     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2718
2719     h->b_stride=  s->mb_width*4;
2720
2721     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2722
2723     s->width = 16*s->mb_width;
2724     s->height= 16*s->mb_height;
2725
2726     if (s->context_initialized
2727         && (   s->width != s->avctx->coded_width || s->height != s->avctx->coded_height
2728             || s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma
2729             || h->cur_chroma_format_idc != h->sps.chroma_format_idc
2730             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2731         if(h != h0) {
2732             av_log_missing_feature(s->avctx, "Width/height/bit depth/chroma idc changing with threads is", 0);
2733             return -1;   // width / height changed during parallelized decoding
2734         }
2735         free_tables(h, 0);
2736         flush_dpb(s->avctx);
2737         MPV_common_end(s);
2738         h->list_count = 0;
2739     }
2740     if (!s->context_initialized) {
2741         if (h != h0) {
2742             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2743             return -1;
2744         }
2745         avcodec_set_dimensions(s->avctx, s->width, s->height);
2746         s->avctx->width  -= (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2747         s->avctx->height -= (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1) * (2 - h->sps.frame_mbs_only_flag);
2748         s->avctx->sample_aspect_ratio= h->sps.sar;
2749         av_assert0(s->avctx->sample_aspect_ratio.den);
2750
2751         if (s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
2752             h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
2753             if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10 &&
2754                 (h->sps.bit_depth_luma != 9 || !CHROMA422)) {
2755                 s->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
2756                 h->cur_chroma_format_idc = h->sps.chroma_format_idc;
2757                 h->pixel_shift = h->sps.bit_depth_luma > 8;
2758
2759                 ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2760                 ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2761                 s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
2762                 dsputil_init(&s->dsp, s->avctx);
2763             } else {
2764                 av_log(s->avctx, AV_LOG_ERROR, "Unsupported bit depth: %d chroma_idc: %d\n",
2765                        h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2766                 return -1;
2767             }
2768         }
2769
2770         if(h->sps.video_signal_type_present_flag){
2771             s->avctx->color_range = h->sps.full_range>0 ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2772             if(h->sps.colour_description_present_flag){
2773                 s->avctx->color_primaries = h->sps.color_primaries;
2774                 s->avctx->color_trc       = h->sps.color_trc;
2775                 s->avctx->colorspace      = h->sps.colorspace;
2776             }
2777         }
2778
2779         if(h->sps.timing_info_present_flag){
2780             int64_t den= h->sps.time_scale;
2781             if(h->x264_build < 44U)
2782                 den *= 2;
2783             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2784                       h->sps.num_units_in_tick, den, 1<<30);
2785         }
2786
2787         switch (h->sps.bit_depth_luma) {
2788             case 9 :
2789                 if (CHROMA444) {
2790                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2791                         s->avctx->pix_fmt = PIX_FMT_GBRP9;
2792                     } else
2793                         s->avctx->pix_fmt = PIX_FMT_YUV444P9;
2794                 } else if (CHROMA422)
2795                     s->avctx->pix_fmt = PIX_FMT_YUV422P9;
2796                 else
2797                     s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2798                 break;
2799             case 10 :
2800                 if (CHROMA444) {
2801                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2802                         s->avctx->pix_fmt = PIX_FMT_GBRP10;
2803                     } else
2804                         s->avctx->pix_fmt = PIX_FMT_YUV444P10;
2805                 } else if (CHROMA422)
2806                     s->avctx->pix_fmt = PIX_FMT_YUV422P10;
2807                 else
2808                     s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2809                 break;
2810             default:
2811                 if (CHROMA444){
2812                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2813                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2814                        s->avctx->pix_fmt = PIX_FMT_GBR24P;
2815                        av_log(h->s.avctx, AV_LOG_DEBUG, "Detected GBR colorspace.\n");
2816                     } else if (s->avctx->colorspace == AVCOL_SPC_YCGCO) {
2817                         av_log(h->s.avctx, AV_LOG_WARNING, "Detected unsupported YCgCo colorspace.\n");
2818                     }
2819                 } else if (CHROMA422) {
2820                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P : PIX_FMT_YUV422P;
2821                 }else{
2822                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2823                                                              s->avctx->codec->pix_fmts ?
2824                                                              s->avctx->codec->pix_fmts :
2825                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2826                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2827                                                              ff_hwaccel_pixfmt_list_420);
2828                 }
2829         }
2830
2831         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2832
2833         if (MPV_common_init(s) < 0) {
2834             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2835             return -1;
2836         }
2837         s->first_field = 0;
2838         h->prev_interlaced_frame = 1;
2839
2840         init_scan_tables(h);
2841         if (ff_h264_alloc_tables(h) < 0) {
2842             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2843             return AVERROR(ENOMEM);
2844         }
2845
2846         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2847             if (context_init(h) < 0) {
2848                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2849                 return -1;
2850             }
2851         } else {
2852             for(i = 1; i < s->avctx->thread_count; i++) {
2853                 H264Context *c;
2854                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2855                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2856                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2857                 c->h264dsp = h->h264dsp;
2858                 c->sps = h->sps;
2859                 c->pps = h->pps;
2860                 c->pixel_shift = h->pixel_shift;
2861                 c->cur_chroma_format_idc = h->cur_chroma_format_idc;
2862                 init_scan_tables(c);
2863                 clone_tables(c, h, i);
2864             }
2865
2866             for(i = 0; i < s->avctx->thread_count; i++)
2867                 if (context_init(h->thread_context[i]) < 0) {
2868                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2869                     return -1;
2870                 }
2871         }
2872     }
2873
2874     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2875
2876     h->mb_mbaff = 0;
2877     h->mb_aff_frame = 0;
2878     last_pic_structure = s0->picture_structure;
2879     if(h->sps.frame_mbs_only_flag){
2880         s->picture_structure= PICT_FRAME;
2881     }else{
2882         if(!h->sps.direct_8x8_inference_flag && slice_type == AV_PICTURE_TYPE_B){
2883             av_log(h->s.avctx, AV_LOG_ERROR, "This stream was generated by a broken encoder, invalid 8x8 inference\n");
2884             return -1;
2885         }
2886         if(get_bits1(&s->gb)) { //field_pic_flag
2887             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2888         } else {
2889             s->picture_structure= PICT_FRAME;
2890             h->mb_aff_frame = h->sps.mb_aff;
2891         }
2892     }
2893     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2894
2895     if(h0->current_slice == 0){
2896         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2897         if(h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0) {
2898             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2899
2900             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2901
2902             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2903                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2904                 if (unwrap_prev_frame_num < 0)
2905                     unwrap_prev_frame_num += max_frame_num;
2906
2907                 h->prev_frame_num = unwrap_prev_frame_num;
2908             }
2909         }
2910
2911         while(h->frame_num !=  h->prev_frame_num && h->prev_frame_num >= 0 &&
2912               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2913             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2914             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2915             if (ff_h264_frame_start(h) < 0)
2916                 return -1;
2917             h->prev_frame_num++;
2918             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2919             s->current_picture_ptr->frame_num= h->prev_frame_num;
2920             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2921             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2922             ff_generate_sliding_window_mmcos(h);
2923             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2924                 (s->avctx->err_recognition & AV_EF_EXPLODE))
2925                 return AVERROR_INVALIDDATA;
2926             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2927              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2928              * about there being no actual duplicates.
2929              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2930              * concealing a lost frame, this probably isn't noticeable by comparison, but it should
2931              * be fixed. */
2932             if (h->short_ref_count) {
2933                 if (prev) {
2934                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2935                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2936                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2937                     h->short_ref[0]->poc = prev->poc+2;
2938                 }
2939                 h->short_ref[0]->frame_num = h->prev_frame_num;
2940             }
2941         }
2942
2943         /* See if we have a decoded first field looking for a pair... */
2944         if (s0->first_field) {
2945             assert(s0->current_picture_ptr);
2946             assert(s0->current_picture_ptr->f.data[0]);
2947             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2948
2949             /* figure out if we have a complementary field pair */
2950             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2951                 /*
2952                  * Previous field is unmatched. Don't display it, but let it
2953                  * remain for reference if marked as such.
2954                  */
2955                 s0->current_picture_ptr = NULL;
2956                 s0->first_field = FIELD_PICTURE;
2957
2958             } else {
2959                 if (s0->current_picture_ptr->frame_num != h->frame_num) {
2960                     /*
2961                      * This and previous field had
2962                      * different frame_nums. Consider this field first in
2963                      * pair. Throw away previous field except for reference
2964                      * purposes.
2965                      */
2966                     s0->first_field = 1;
2967                     s0->current_picture_ptr = NULL;
2968
2969                 } else {
2970                     /* Second field in complementary pair */
2971                     s0->first_field = 0;
2972                 }
2973             }
2974
2975         } else {
2976             /* Frame or first field in a potentially complementary pair */
2977             assert(!s0->current_picture_ptr);
2978             s0->first_field = FIELD_PICTURE;
2979         }
2980
2981         if(!FIELD_PICTURE || s0->first_field) {
2982             if (ff_h264_frame_start(h) < 0) {
2983                 s0->first_field = 0;
2984                 return -1;
2985             }
2986         } else {
2987             ff_release_unused_pictures(s, 0);
2988         }
2989     }
2990     if(h != h0)
2991         clone_slice(h, h0);
2992
2993     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2994
2995     assert(s->mb_num == s->mb_width * s->mb_height);
2996     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2997        first_mb_in_slice                    >= s->mb_num){
2998         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2999         return -1;
3000     }
3001     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
3002     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
3003     if (s->picture_structure == PICT_BOTTOM_FIELD)
3004         s->resync_mb_y = s->mb_y = s->mb_y + 1;
3005     assert(s->mb_y < s->mb_height);
3006
3007     if(s->picture_structure==PICT_FRAME){
3008         h->curr_pic_num=   h->frame_num;
3009         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
3010     }else{
3011         h->curr_pic_num= 2*h->frame_num + 1;
3012         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
3013     }
3014
3015     if(h->nal_unit_type == NAL_IDR_SLICE){
3016         get_ue_golomb(&s->gb); /* idr_pic_id */
3017     }
3018
3019     if(h->sps.poc_type==0){
3020         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
3021
3022         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
3023             h->delta_poc_bottom= get_se_golomb(&s->gb);
3024         }
3025     }
3026
3027     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
3028         h->delta_poc[0]= get_se_golomb(&s->gb);
3029
3030         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
3031             h->delta_poc[1]= get_se_golomb(&s->gb);
3032     }
3033
3034     init_poc(h);
3035
3036     if(h->pps.redundant_pic_cnt_present){
3037         h->redundant_pic_count= get_ue_golomb(&s->gb);
3038     }
3039
3040     //set defaults, might be overridden a few lines later
3041     h->ref_count[0]= h->pps.ref_count[0];
3042     h->ref_count[1]= h->pps.ref_count[1];
3043
3044     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
3045         unsigned max= (16<<(s->picture_structure != PICT_FRAME))-1;
3046         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
3047             h->direct_spatial_mv_pred= get_bits1(&s->gb);
3048         }
3049         num_ref_idx_active_override_flag= get_bits1(&s->gb);
3050
3051         if(num_ref_idx_active_override_flag){
3052             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
3053             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
3054                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
3055
3056         }
3057         if(h->ref_count[0]-1 > max || h->ref_count[1]-1 > max){
3058             av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3059             h->ref_count[0]= h->ref_count[1]= 1;
3060             return -1;
3061         }
3062         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
3063             h->list_count= 2;
3064         else
3065             h->list_count= 1;
3066     }else
3067         h->ref_count[1]= h->ref_count[0]= h->list_count= 0;
3068
3069     if(!default_ref_list_done){
3070         ff_h264_fill_default_ref_list(h);
3071     }
3072
3073     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0) {
3074         h->ref_count[1]= h->ref_count[0]= 0;
3075         return -1;
3076     }
3077
3078     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
3079         s->last_picture_ptr= &h->ref_list[0][0];
3080         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
3081     }
3082     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
3083         s->next_picture_ptr= &h->ref_list[1][0];
3084         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
3085     }
3086
3087     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
3088        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
3089         pred_weight_table(h);
3090     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
3091         implicit_weight_table(h, -1);
3092     }else {
3093         h->use_weight = 0;
3094         for (i = 0; i < 2; i++) {
3095             h->luma_weight_flag[i]   = 0;
3096             h->chroma_weight_flag[i] = 0;
3097         }
3098     }
3099
3100     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
3101        (s->avctx->err_recognition & AV_EF_EXPLODE))
3102         return AVERROR_INVALIDDATA;
3103
3104     if(FRAME_MBAFF){
3105         ff_h264_fill_mbaff_ref_list(h);
3106
3107         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
3108             implicit_weight_table(h, 0);
3109             implicit_weight_table(h, 1);
3110         }
3111     }
3112
3113     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
3114         ff_h264_direct_dist_scale_factor(h);
3115     ff_h264_direct_ref_list_init(h);
3116
3117     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
3118         tmp = get_ue_golomb_31(&s->gb);
3119         if(tmp > 2){
3120             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3121             return -1;
3122         }
3123         h->cabac_init_idc= tmp;
3124     }
3125
3126     h->last_qscale_diff = 0;
3127     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3128     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
3129         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3130         return -1;
3131     }
3132     s->qscale= tmp;
3133     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3134     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3135     //FIXME qscale / qp ... stuff
3136     if(h->slice_type == AV_PICTURE_TYPE_SP){
3137         get_bits1(&s->gb); /* sp_for_switch_flag */
3138     }
3139     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
3140         get_se_golomb(&s->gb); /* slice_qs_delta */
3141     }
3142
3143     h->deblocking_filter = 1;
3144     h->slice_alpha_c0_offset = 52;
3145     h->slice_beta_offset = 52;
3146     if( h->pps.deblocking_filter_parameters_present ) {
3147         tmp= get_ue_golomb_31(&s->gb);
3148         if(tmp > 2){
3149             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
3150             return -1;
3151         }
3152         h->deblocking_filter= tmp;
3153         if(h->deblocking_filter < 2)
3154             h->deblocking_filter^= 1; // 1<->0
3155
3156         if( h->deblocking_filter ) {
3157             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3158             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
3159             if(   h->slice_alpha_c0_offset > 104U
3160                || h->slice_beta_offset     > 104U){
3161                 av_log(s->avctx, AV_LOG_ERROR, "deblocking filter parameters %d %d out of range\n", h->slice_alpha_c0_offset, h->slice_beta_offset);
3162                 return -1;
3163             }
3164         }
3165     }
3166
3167     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
3168        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
3169        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
3170        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
3171         h->deblocking_filter= 0;
3172
3173     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
3174         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
3175             /* Cheat slightly for speed:
3176                Do not bother to deblock across slices. */
3177             h->deblocking_filter = 2;
3178         } else {
3179             h0->max_contexts = 1;
3180             if(!h0->single_decode_warning) {
3181                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3182                 h0->single_decode_warning = 1;
3183             }
3184             if (h != h0) {
3185                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
3186                 return 1;
3187             }
3188         }
3189     }
3190     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
3191                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
3192                  + 6 * (h->sps.bit_depth_luma - 8);
3193
3194 #if 0 //FMO
3195     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
3196         slice_group_change_cycle= get_bits(&s->gb, ?);
3197 #endif
3198
3199     h0->last_slice_type = slice_type;
3200     h->slice_num = ++h0->current_slice;
3201
3202     if(h->slice_num)
3203         h0->slice_row[(h->slice_num-1)&(MAX_SLICES-1)]= s->resync_mb_y;
3204     if (   h0->slice_row[h->slice_num&(MAX_SLICES-1)] + 3 >= s->resync_mb_y
3205         && h0->slice_row[h->slice_num&(MAX_SLICES-1)] <= s->resync_mb_y
3206         && h->slice_num >= MAX_SLICES) {
3207         //in case of ASO this check needs to be updated depending on how we decide to assign slice numbers in this case
3208         av_log(s->avctx, AV_LOG_WARNING, "Possibly too many slices (%d >= %d), increase MAX_SLICES and recompile if there are artifacts\n", h->slice_num, MAX_SLICES);
3209     }
3210
3211     for(j=0; j<2; j++){
3212         int id_list[16];
3213         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
3214         for(i=0; i<16; i++){
3215             id_list[i]= 60;
3216             if (h->ref_list[j][i].f.data[0]) {
3217                 int k;
3218                 uint8_t *base = h->ref_list[j][i].f.base[0];
3219                 for(k=0; k<h->short_ref_count; k++)
3220                     if (h->short_ref[k]->f.base[0] == base) {
3221                         id_list[i]= k;
3222                         break;
3223                     }
3224                 for(k=0; k<h->long_ref_count; k++)
3225                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3226                         id_list[i]= h->short_ref_count + k;
3227                         break;
3228                     }
3229             }
3230         }
3231
3232         ref2frm[0]=
3233         ref2frm[1]= -1;
3234         for(i=0; i<16; i++)
3235             ref2frm[i+2]= 4*id_list[i]
3236                           + (h->ref_list[j][i].f.reference & 3);
3237         ref2frm[18+0]=
3238         ref2frm[18+1]= -1;
3239         for(i=16; i<48; i++)
3240             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3241                           + (h->ref_list[j][i].f.reference & 3);
3242     }
3243
3244     //FIXME: fix draw_edges+PAFF+frame threads
3245     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3246     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3247
3248     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3249         av_log(h->s.avctx, AV_LOG_DEBUG, "slice:%d %s mb:%d %c%s%s pps:%u frame:%d poc:%d/%d ref:%d/%d qp:%d loop:%d:%d:%d weight:%d%s %s\n",
3250                h->slice_num,
3251                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3252                first_mb_in_slice,
3253                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3254                pps_id, h->frame_num,
3255                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3256                h->ref_count[0], h->ref_count[1],
3257                s->qscale,
3258                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3259                h->use_weight,
3260                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3261                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3262                );
3263     }
3264
3265     return 0;
3266 }
3267
3268 int ff_h264_get_slice_type(const H264Context *h)
3269 {
3270     switch (h->slice_type) {
3271     case AV_PICTURE_TYPE_P:  return 0;
3272     case AV_PICTURE_TYPE_B:  return 1;
3273     case AV_PICTURE_TYPE_I:  return 2;
3274     case AV_PICTURE_TYPE_SP: return 3;
3275     case AV_PICTURE_TYPE_SI: return 4;
3276     default:         return -1;
3277     }
3278 }
3279
3280 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3281                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3282 {
3283     int b_stride = h->b_stride;
3284     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3285     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3286     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3287         if(USES_LIST(top_type, list)){
3288             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3289             const int b8_xy= 4*top_xy + 2;
3290             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3291             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3292             ref_cache[0 - 1*8]=
3293             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3294             ref_cache[2 - 1*8]=
3295             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3296         }else{
3297             AV_ZERO128(mv_dst - 1*8);
3298             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3299         }
3300
3301         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3302             if(USES_LIST(left_type[LTOP], list)){
3303                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3304                 const int b8_xy= 4*left_xy[LTOP] + 1;
3305                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3306                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3307                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3308                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3309                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3310                 ref_cache[-1 +  0]=
3311                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3312                 ref_cache[-1 + 16]=
3313                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3314             }else{
3315                 AV_ZERO32(mv_dst - 1 + 0);
3316                 AV_ZERO32(mv_dst - 1 + 8);
3317                 AV_ZERO32(mv_dst - 1 +16);
3318                 AV_ZERO32(mv_dst - 1 +24);
3319                 ref_cache[-1 +  0]=
3320                 ref_cache[-1 +  8]=
3321                 ref_cache[-1 + 16]=
3322                 ref_cache[-1 + 24]= LIST_NOT_USED;
3323             }
3324         }
3325     }
3326
3327     if(!USES_LIST(mb_type, list)){
3328         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3329         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3330         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3331         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3332         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3333         return;
3334     }
3335
3336     {
3337         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3338         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3339         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3340         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3341         AV_WN32A(&ref_cache[0*8], ref01);
3342         AV_WN32A(&ref_cache[1*8], ref01);
3343         AV_WN32A(&ref_cache[2*8], ref23);
3344         AV_WN32A(&ref_cache[3*8], ref23);
3345     }
3346
3347     {
3348         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3349         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3350         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3351         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3352         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3353     }
3354 }
3355
3356 /**
3357  *
3358  * @return non zero if the loop filter can be skipped
3359  */
3360 static int fill_filter_caches(H264Context *h, int mb_type){
3361     MpegEncContext * const s = &h->s;
3362     const int mb_xy= h->mb_xy;
3363     int top_xy, left_xy[LEFT_MBS];
3364     int top_type, left_type[LEFT_MBS];
3365     uint8_t *nnz;
3366     uint8_t *nnz_cache;
3367
3368     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3369
3370     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3371      * stuff, I can't imagine that these complex rules are worth it. */
3372
3373     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3374     if(FRAME_MBAFF){
3375         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3376         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3377         if(s->mb_y&1){
3378             if (left_mb_field_flag != curr_mb_field_flag) {
3379                 left_xy[LTOP] -= s->mb_stride;
3380             }
3381         }else{
3382             if(curr_mb_field_flag){
3383                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3384             }
3385             if (left_mb_field_flag != curr_mb_field_flag) {
3386                 left_xy[LBOT] += s->mb_stride;
3387             }
3388         }
3389     }
3390
3391     h->top_mb_xy = top_xy;
3392     h->left_mb_xy[LTOP] = left_xy[LTOP];
3393     h->left_mb_xy[LBOT] = left_xy[LBOT];
3394     {
3395         //for sufficiently low qp, filtering wouldn't do anything
3396         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3397         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3398         int qp = s->current_picture.f.qscale_table[mb_xy];
3399         if(qp <= qp_thresh
3400            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3401            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3402             if(!FRAME_MBAFF)
3403                 return 1;
3404             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3405                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3406                 return 1;
3407         }
3408     }
3409
3410     top_type        = s->current_picture.f.mb_type[top_xy];
3411     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3412     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3413     if(h->deblocking_filter == 2){
3414         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3415         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3416     }else{
3417         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3418         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3419     }
3420     h->top_type       = top_type;
3421     h->left_type[LTOP]= left_type[LTOP];
3422     h->left_type[LBOT]= left_type[LBOT];
3423
3424     if(IS_INTRA(mb_type))
3425         return 0;
3426
3427     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3428     if(h->list_count == 2)
3429         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3430
3431     nnz = h->non_zero_count[mb_xy];
3432     nnz_cache = h->non_zero_count_cache;
3433     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3434     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3435     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3436     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3437     h->cbp= h->cbp_table[mb_xy];
3438
3439     if(top_type){
3440         nnz = h->non_zero_count[top_xy];
3441         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3442     }
3443
3444     if(left_type[LTOP]){
3445         nnz = h->non_zero_count[left_xy[LTOP]];
3446         nnz_cache[3+8*1]= nnz[3+0*4];
3447         nnz_cache[3+8*2]= nnz[3+1*4];
3448         nnz_cache[3+8*3]= nnz[3+2*4];
3449         nnz_cache[3+8*4]= nnz[3+3*4];
3450     }
3451
3452     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3453     if(!CABAC && h->pps.transform_8x8_mode){
3454         if(IS_8x8DCT(top_type)){
3455             nnz_cache[4+8*0]=
3456             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3457             nnz_cache[6+8*0]=
3458             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3459         }
3460         if(IS_8x8DCT(left_type[LTOP])){
3461             nnz_cache[3+8*1]=
3462             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3463         }
3464         if(IS_8x8DCT(left_type[LBOT])){
3465             nnz_cache[3+8*3]=
3466             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3467         }
3468
3469         if(IS_8x8DCT(mb_type)){
3470             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3471             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3472
3473             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3474             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3475
3476             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3477             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3478
3479             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3480             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3481         }
3482     }
3483
3484     return 0;
3485 }
3486
3487 static void loop_filter(H264Context *h, int start_x, int end_x){
3488     MpegEncContext * const s = &h->s;
3489     uint8_t  *dest_y, *dest_cb, *dest_cr;
3490     int linesize, uvlinesize, mb_x, mb_y;
3491     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3492     const int old_slice_type= h->slice_type;
3493     const int pixel_shift = h->pixel_shift;
3494     const int block_h = 16 >> s->chroma_y_shift;
3495
3496     if(h->deblocking_filter) {
3497         for(mb_x= start_x; mb_x<end_x; mb_x++){
3498             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3499                 int mb_xy, mb_type;
3500                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3501                 h->slice_num= h->slice_table[mb_xy];
3502                 mb_type = s->current_picture.f.mb_type[mb_xy];
3503                 h->list_count= h->list_counts[mb_xy];
3504
3505                 if(FRAME_MBAFF)
3506                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3507
3508                 s->mb_x= mb_x;
3509                 s->mb_y= mb_y;
3510                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3511                 dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3512                 dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3513                     //FIXME simplify above
3514
3515                 if (MB_FIELD) {
3516                     linesize   = h->mb_linesize   = s->linesize * 2;
3517                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3518                     if(mb_y&1){ //FIXME move out of this function?
3519                         dest_y -= s->linesize*15;
3520                         dest_cb-= s->uvlinesize * (block_h - 1);
3521                         dest_cr-= s->uvlinesize * (block_h - 1);
3522                     }
3523                 } else {
3524                     linesize   = h->mb_linesize   = s->linesize;
3525                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3526                 }
3527                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
3528                 if(fill_filter_caches(h, mb_type))
3529                     continue;
3530                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3531                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3532
3533                 if (FRAME_MBAFF) {
3534                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3535                 } else {
3536                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3537                 }
3538             }
3539         }
3540     }
3541     h->slice_type= old_slice_type;
3542     s->mb_x= end_x;
3543     s->mb_y= end_mb_y - FRAME_MBAFF;
3544     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3545     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3546 }
3547
3548 static void predict_field_decoding_flag(H264Context *h){
3549     MpegEncContext * const s = &h->s;
3550     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3551     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3552                 ? s->current_picture.f.mb_type[mb_xy - 1]
3553                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3554                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3555                 : 0;
3556     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3557 }
3558
3559 /**
3560  * Draw edges and report progress for the last MB row.
3561  */
3562 static void decode_finish_row(H264Context *h){
3563     MpegEncContext * const s = &h->s;
3564     int top = 16*(s->mb_y >> FIELD_PICTURE);
3565     int height = 16 << FRAME_MBAFF;
3566     int deblock_border = (16 + 4) << FRAME_MBAFF;
3567     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3568
3569     if (h->deblocking_filter) {
3570         if((top + height) >= pic_height)
3571             height += deblock_border;
3572
3573         top -= deblock_border;
3574     }
3575
3576     if (top >= pic_height || (top + height) < h->emu_edge_height)
3577         return;
3578
3579     height = FFMIN(height, pic_height - top);
3580     if (top < h->emu_edge_height) {
3581         height = top+height;
3582         top = 0;
3583     }
3584
3585     ff_draw_horiz_band(s, top, height);
3586
3587     if (s->dropable) return;
3588
3589     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3590                              s->picture_structure==PICT_BOTTOM_FIELD);
3591 }
3592
3593 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3594     H264Context *h = *(void**)arg;
3595     MpegEncContext * const s = &h->s;
3596     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3597     int lf_x_start = s->mb_x;
3598
3599     s->mb_skip_run= -1;
3600
3601     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3602                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3603
3604     if( h->pps.cabac ) {
3605         /* realign */
3606         align_get_bits( &s->gb );
3607
3608         /* init cabac */
3609         ff_init_cabac_states( &h->cabac);
3610         ff_init_cabac_decoder( &h->cabac,
3611                                s->gb.buffer + get_bits_count(&s->gb)/8,
3612                                (get_bits_left(&s->gb) + 7)/8);
3613
3614         ff_h264_init_cabac_states(h);
3615
3616         for(;;){
3617 //START_TIMER
3618             int ret = ff_h264_decode_mb_cabac(h);
3619             int eos;
3620 //STOP_TIMER("decode_mb_cabac")
3621
3622             if(ret>=0) ff_h264_hl_decode_mb(h);
3623
3624             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3625                 s->mb_y++;
3626
3627                 ret = ff_h264_decode_mb_cabac(h);
3628
3629                 if(ret>=0) ff_h264_hl_decode_mb(h);
3630                 s->mb_y--;
3631             }
3632             eos = get_cabac_terminate( &h->cabac );
3633
3634             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3635                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3636                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3637                 return 0;
3638             }
3639             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3640                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d, bytestream (%td)\n", s->mb_x, s->mb_y, h->cabac.bytestream_end - h->cabac.bytestream);
3641                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3642                 return -1;
3643             }
3644
3645             if( ++s->mb_x >= s->mb_width ) {
3646                 loop_filter(h, lf_x_start, s->mb_x);
3647                 s->mb_x = lf_x_start = 0;
3648                 decode_finish_row(h);
3649                 ++s->mb_y;
3650                 if(FIELD_OR_MBAFF_PICTURE) {
3651                     ++s->mb_y;
3652                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3653                         predict_field_decoding_flag(h);
3654                 }
3655             }
3656
3657             if( eos || s->mb_y >= s->mb_height ) {
3658                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3659                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3660                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3661                 return 0;
3662             }
3663         }
3664
3665     } else {
3666         for(;;){
3667             int ret = ff_h264_decode_mb_cavlc(h);
3668
3669             if(ret>=0) ff_h264_hl_decode_mb(h);
3670
3671             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3672                 s->mb_y++;
3673                 ret = ff_h264_decode_mb_cavlc(h);
3674
3675                 if(ret>=0) ff_h264_hl_decode_mb(h);
3676                 s->mb_y--;
3677             }
3678
3679             if(ret<0){
3680                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3681                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3682                 return -1;
3683             }
3684
3685             if(++s->mb_x >= s->mb_width){
3686                 loop_filter(h, lf_x_start, s->mb_x);
3687                 s->mb_x = lf_x_start = 0;
3688                 decode_finish_row(h);
3689                 ++s->mb_y;
3690                 if(FIELD_OR_MBAFF_PICTURE) {
3691                     ++s->mb_y;
3692                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3693                         predict_field_decoding_flag(h);
3694                 }
3695                 if(s->mb_y >= s->mb_height){
3696                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3697
3698                     if(   get_bits_count(&s->gb) == s->gb.size_in_bits
3699                        || get_bits_count(&s->gb) <  s->gb.size_in_bits && s->avctx->error_recognition < FF_ER_AGGRESSIVE) {
3700                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3701
3702                         return 0;
3703                     }else{
3704                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3705
3706                         return -1;
3707                     }
3708                 }
3709             }
3710
3711             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3712                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3713                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3714                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3715                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3716
3717                     return 0;
3718                 }else{
3719                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3720
3721                     return -1;
3722                 }
3723             }
3724         }
3725     }
3726 }
3727
3728 /**
3729  * Call decode_slice() for each context.
3730  *
3731  * @param h h264 master context
3732  * @param context_count number of contexts to execute
3733  */
3734 static int execute_decode_slices(H264Context *h, int context_count){
3735     MpegEncContext * const s = &h->s;
3736     AVCodecContext * const avctx= s->avctx;
3737     H264Context *hx;
3738     int i;
3739
3740     if (s->avctx->hwaccel || s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3741         return 0;
3742     if(context_count == 1) {
3743         return decode_slice(avctx, &h);
3744     } else {
3745         for(i = 1; i < context_count; i++) {
3746             hx = h->thread_context[i];
3747             hx->s.err_recognition = avctx->err_recognition;
3748             hx->s.error_count = 0;
3749             hx->x264_build= h->x264_build;
3750         }
3751
3752         avctx->execute(avctx, decode_slice,
3753                        h->thread_context, NULL, context_count, sizeof(void*));
3754
3755         /* pull back stuff from slices to master context */
3756         hx = h->thread_context[context_count - 1];
3757         s->mb_x = hx->s.mb_x;
3758         s->mb_y = hx->s.mb_y;
3759         s->dropable = hx->s.dropable;
3760         s->picture_structure = hx->s.picture_structure;
3761         for(i = 1; i < context_count; i++)
3762             h->s.error_count += h->thread_context[i]->s.error_count;
3763     }
3764
3765     return 0;
3766 }
3767
3768
3769 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3770     MpegEncContext * const s = &h->s;
3771     AVCodecContext * const avctx= s->avctx;
3772     H264Context *hx; ///< thread context
3773     int buf_index;
3774     int context_count;
3775     int next_avc;
3776     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3777     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3778     int nal_index;
3779
3780     h->max_contexts = (HAVE_THREADS && (s->avctx->active_thread_type&FF_THREAD_SLICE)) ? avctx->thread_count : 1;
3781     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3782         h->current_slice = 0;
3783         if (!s->first_field)
3784             s->current_picture_ptr= NULL;
3785         ff_h264_reset_sei(h);
3786     }
3787
3788     for(;pass <= 1;pass++){
3789         buf_index = 0;
3790         context_count = 0;
3791         next_avc = h->is_avc ? 0 : buf_size;
3792         nal_index = 0;
3793     for(;;){
3794         int consumed;
3795         int dst_length;
3796         int bit_length;
3797         const uint8_t *ptr;
3798         int i, nalsize = 0;
3799         int err;
3800
3801         if(buf_index >= next_avc) {
3802             if(buf_index >= buf_size) break;
3803             nalsize = 0;
3804             for(i = 0; i < h->nal_length_size; i++)
3805                 nalsize = (nalsize << 8) | buf[buf_index++];
3806             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3807                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3808                 break;
3809             }
3810             next_avc= buf_index + nalsize;
3811         } else {
3812             // start code prefix search
3813             for(; buf_index + 3 < next_avc; buf_index++){
3814                 // This should always succeed in the first iteration.
3815                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3816                     break;
3817             }
3818
3819             if(buf_index+3 >= buf_size) break;
3820
3821             buf_index+=3;
3822             if(buf_index >= next_avc) continue;
3823         }
3824
3825         hx = h->thread_context[context_count];
3826
3827         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
3828         if (ptr==NULL || dst_length < 0){
3829             return -1;
3830         }
3831         i= buf_index + consumed;
3832         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3833            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3834             s->workaround_bugs |= FF_BUG_TRUNCATED;
3835
3836         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3837         while(dst_length > 0 && ptr[dst_length - 1] == 0)
3838             dst_length--;
3839         }
3840         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3841
3842         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3843             av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d/%d at %d/%d length %d\n", hx->nal_unit_type, hx->nal_ref_idc, buf_index, buf_size, dst_length);
3844         }
3845
3846         if (h->is_avc && (nalsize != consumed) && nalsize){
3847             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3848         }
3849
3850         buf_index += consumed;
3851         nal_index++;
3852
3853         if(pass == 0) {
3854             // packets can sometimes contain multiple PPS/SPS
3855             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3856             // if so, when frame threading we can't start the next thread until we've read all of them
3857             switch (hx->nal_unit_type) {
3858                 case NAL_SPS:
3859                 case NAL_PPS:
3860                     nals_needed = nal_index;
3861                     break;
3862                 case NAL_IDR_SLICE:
3863                 case NAL_SLICE:
3864                     init_get_bits(&hx->s.gb, ptr, bit_length);
3865                     if (!get_ue_golomb(&hx->s.gb))
3866                         nals_needed = nal_index;
3867             }
3868             continue;
3869         }
3870
3871         //FIXME do not discard SEI id
3872         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3873             continue;
3874
3875       again:
3876         err = 0;
3877         switch(hx->nal_unit_type){
3878         case NAL_IDR_SLICE:
3879             if (h->nal_unit_type != NAL_IDR_SLICE) {
3880                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3881                 return -1;
3882             }
3883             idr(h); // FIXME ensure we don't lose some frames if there is reordering
3884         case NAL_SLICE:
3885             init_get_bits(&hx->s.gb, ptr, bit_length);
3886             hx->intra_gb_ptr=
3887             hx->inter_gb_ptr= &hx->s.gb;
3888             hx->s.data_partitioning = 0;
3889
3890             if((err = decode_slice_header(hx, h)))
3891                break;
3892
3893             if (   h->sei_recovery_frame_cnt >= 0
3894                 && ((h->recovery_frame - h->frame_num) & ((1 << h->sps.log2_max_frame_num)-1)) > h->sei_recovery_frame_cnt) {
3895                 h->recovery_frame = (h->frame_num + h->sei_recovery_frame_cnt) %
3896                                     (1 << h->sps.log2_max_frame_num);
3897             }
3898
3899             s->current_picture_ptr->f.key_frame |=
3900                     (hx->nal_unit_type == NAL_IDR_SLICE);
3901
3902             if (h->recovery_frame == h->frame_num) {
3903                 h->sync |= 1;
3904                 h->recovery_frame = -1;
3905             }
3906
3907             h->sync |= !!s->current_picture_ptr->f.key_frame;
3908             h->sync |= 3*!!(s->flags2 & CODEC_FLAG2_SHOW_ALL);
3909             s->current_picture_ptr->sync = h->sync;
3910
3911             if (h->current_slice == 1) {
3912                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3913                     decode_postinit(h, nal_index >= nals_needed);
3914                 }
3915
3916                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3917                     return -1;
3918                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3919                     ff_vdpau_h264_picture_start(s);
3920             }
3921
3922             if(hx->redundant_pic_count==0
3923                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3924                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3925                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3926                && avctx->skip_frame < AVDISCARD_ALL){
3927                 if(avctx->hwaccel) {
3928                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3929                         return -1;
3930                 }else
3931                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3932                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3933                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3934                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3935                 }else
3936                     context_count++;
3937             }
3938             break;
3939         case NAL_DPA:
3940             init_get_bits(&hx->s.gb, ptr, bit_length);
3941             hx->intra_gb_ptr=
3942             hx->inter_gb_ptr= NULL;
3943
3944             if ((err = decode_slice_header(hx, h)) < 0)
3945                 break;
3946
3947             hx->s.data_partitioning = 1;
3948
3949             break;
3950         case NAL_DPB:
3951             init_get_bits(&hx->intra_gb, ptr, bit_length);
3952             hx->intra_gb_ptr= &hx->intra_gb;
3953             break;
3954         case NAL_DPC:
3955             init_get_bits(&hx->inter_gb, ptr, bit_length);
3956             hx->inter_gb_ptr= &hx->inter_gb;
3957
3958             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3959                && s->context_initialized
3960                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3961                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3962                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3963                && avctx->skip_frame < AVDISCARD_ALL)
3964                 context_count++;
3965             break;
3966         case NAL_SEI:
3967             init_get_bits(&s->gb, ptr, bit_length);
3968             ff_h264_decode_sei(h);
3969             break;
3970         case NAL_SPS:
3971             init_get_bits(&s->gb, ptr, bit_length);
3972             if(ff_h264_decode_seq_parameter_set(h) < 0 && (h->is_avc ? (nalsize != consumed) && nalsize : 1)){
3973                 av_log(h->s.avctx, AV_LOG_DEBUG, "SPS decoding failure, trying alternative mode\n");
3974                 if(h->is_avc) av_assert0(next_avc - buf_index + consumed == nalsize);
3975                 init_get_bits(&s->gb, &buf[buf_index + 1 - consumed], 8*(next_avc - buf_index + consumed));
3976                 ff_h264_decode_seq_parameter_set(h);
3977             }
3978
3979             if (s->flags& CODEC_FLAG_LOW_DELAY ||
3980                 (h->sps.bitstream_restriction_flag && !h->sps.num_reorder_frames))
3981                 s->low_delay=1;
3982
3983             if(avctx->has_b_frames < 2)
3984                 avctx->has_b_frames= !s->low_delay;
3985             break;
3986         case NAL_PPS:
3987             init_get_bits(&s->gb, ptr, bit_length);
3988
3989             ff_h264_decode_picture_parameter_set(h, bit_length);
3990
3991             break;
3992         case NAL_AUD:
3993         case NAL_END_SEQUENCE:
3994         case NAL_END_STREAM:
3995         case NAL_FILLER_DATA:
3996         case NAL_SPS_EXT:
3997         case NAL_AUXILIARY_SLICE:
3998             break;
3999         default:
4000             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
4001         }
4002
4003         if(context_count == h->max_contexts) {
4004             execute_decode_slices(h, context_count);
4005             context_count = 0;
4006         }
4007
4008         if (err < 0)
4009             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
4010         else if(err == 1) {
4011             /* Slice could not be decoded in parallel mode, copy down
4012              * NAL unit stuff to context 0 and restart. Note that
4013              * rbsp_buffer is not transferred, but since we no longer
4014              * run in parallel mode this should not be an issue. */
4015             h->nal_unit_type = hx->nal_unit_type;
4016             h->nal_ref_idc   = hx->nal_ref_idc;
4017             hx = h;
4018             goto again;
4019         }
4020     }
4021     }
4022     if(context_count)
4023         execute_decode_slices(h, context_count);
4024     return buf_index;
4025 }
4026
4027 /**
4028  * Return the number of bytes consumed for building the current frame.
4029  */
4030 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
4031         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
4032         if(pos+10>buf_size) pos=buf_size; // oops ;)
4033
4034         return pos;
4035 }
4036
4037 static int decode_frame(AVCodecContext *avctx,
4038                              void *data, int *data_size,
4039                              AVPacket *avpkt)
4040 {
4041     const uint8_t *buf = avpkt->data;
4042     int buf_size = avpkt->size;
4043     H264Context *h = avctx->priv_data;
4044     MpegEncContext *s = &h->s;
4045     AVFrame *pict = data;
4046     int buf_index;
4047
4048     s->flags= avctx->flags;
4049     s->flags2= avctx->flags2;
4050
4051    /* end of stream, output what is still in the buffers */
4052  out:
4053     if (buf_size == 0) {
4054         Picture *out;
4055         int i, out_idx;
4056
4057         s->current_picture_ptr = NULL;
4058
4059 //FIXME factorize this with the output code below
4060         out = h->delayed_pic[0];
4061         out_idx = 0;
4062         for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
4063             if(h->delayed_pic[i]->poc < out->poc){
4064                 out = h->delayed_pic[i];
4065                 out_idx = i;
4066             }
4067
4068         for(i=out_idx; h->delayed_pic[i]; i++)
4069             h->delayed_pic[i] = h->delayed_pic[i+1];
4070
4071         if(out){
4072             *data_size = sizeof(AVFrame);
4073             *pict= *(AVFrame*)out;
4074         }
4075
4076         return 0;
4077     }
4078     if(h->is_avc && buf_size >= 9 && AV_RB32(buf)==0x0164001F && buf[5] && buf[8]==0x67)
4079         return ff_h264_decode_extradata(h, buf, buf_size);
4080
4081     buf_index=decode_nal_units(h, buf, buf_size);
4082     if(buf_index < 0)
4083         return -1;
4084
4085     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
4086         buf_size = 0;
4087         goto out;
4088     }
4089
4090     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
4091         if (avctx->skip_frame >= AVDISCARD_NONREF ||
4092             buf_size >= 4 && !memcmp("Q264", buf, 4))
4093             return 0;
4094         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
4095         return -1;
4096     }
4097
4098     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
4099
4100         if(s->flags2 & CODEC_FLAG2_CHUNKS) decode_postinit(h, 1);
4101
4102         field_end(h, 0);
4103
4104         *data_size = 0; /* Wait for second field. */
4105         if (h->next_output_pic && h->next_output_pic->sync) {
4106             if(h->sync>1 || h->next_output_pic->f.pict_type != AV_PICTURE_TYPE_B){
4107                 *data_size = sizeof(AVFrame);
4108                 *pict = *(AVFrame*)h->next_output_pic;
4109             }
4110         }
4111     }
4112
4113     assert(pict->data[0] || !*data_size);
4114     ff_print_debug_info(s, pict);
4115 //printf("out %d\n", (int)pict->data[0]);
4116
4117     return get_consumed_bytes(s, buf_index, buf_size);
4118 }
4119 #if 0
4120 static inline void fill_mb_avail(H264Context *h){
4121     MpegEncContext * const s = &h->s;
4122     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
4123
4124     if(s->mb_y){
4125         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
4126         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
4127         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
4128     }else{
4129         h->mb_avail[0]=
4130         h->mb_avail[1]=
4131         h->mb_avail[2]= 0;
4132     }
4133     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
4134     h->mb_avail[4]= 1; //FIXME move out
4135     h->mb_avail[5]= 0; //FIXME move out
4136 }
4137 #endif
4138
4139 #ifdef TEST
4140 #undef printf
4141 #undef random
4142 #define COUNT 8000
4143 #define SIZE (COUNT*40)
4144 extern AVCodec ff_h264_decoder;
4145 int main(void){
4146     int i;
4147     uint8_t temp[SIZE];
4148     PutBitContext pb;
4149     GetBitContext gb;
4150 //    int int_temp[10000];
4151     DSPContext dsp;
4152     AVCodecContext avctx;
4153
4154     avcodec_get_context_defaults3(&avctx, &ff_h264_decoder);
4155
4156     dsputil_init(&dsp, &avctx);
4157
4158     init_put_bits(&pb, temp, SIZE);
4159     printf("testing unsigned exp golomb\n");
4160     for(i=0; i<COUNT; i++){
4161         START_TIMER
4162         set_ue_golomb(&pb, i);
4163         STOP_TIMER("set_ue_golomb");
4164     }
4165     flush_put_bits(&pb);
4166
4167     init_get_bits(&gb, temp, 8*SIZE);
4168     for(i=0; i<COUNT; i++){
4169         int j, s;
4170
4171         s= show_bits(&gb, 24);
4172
4173         START_TIMER
4174         j= get_ue_golomb(&gb);
4175         if(j != i){
4176             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
4177 //            return -1;
4178         }
4179         STOP_TIMER("get_ue_golomb");
4180     }
4181
4182
4183     init_put_bits(&pb, temp, SIZE);
4184     printf("testing signed exp golomb\n");
4185     for(i=0; i<COUNT; i++){
4186         START_TIMER
4187         set_se_golomb(&pb, i - COUNT/2);
4188         STOP_TIMER("set_se_golomb");
4189     }
4190     flush_put_bits(&pb);
4191
4192     init_get_bits(&gb, temp, 8*SIZE);
4193     for(i=0; i<COUNT; i++){
4194         int j, s;
4195
4196         s= show_bits(&gb, 24);
4197
4198         START_TIMER
4199         j= get_se_golomb(&gb);
4200         if(j != i - COUNT/2){
4201             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
4202 //            return -1;
4203         }
4204         STOP_TIMER("get_se_golomb");
4205     }
4206
4207     printf("Testing RBSP\n");
4208
4209
4210     return 0;
4211 }
4212 #endif /* TEST */
4213
4214
4215 av_cold void ff_h264_free_context(H264Context *h)
4216 {
4217     int i;
4218
4219     free_tables(h, 1); //FIXME cleanup init stuff perhaps
4220
4221     for(i = 0; i < MAX_SPS_COUNT; i++)
4222         av_freep(h->sps_buffers + i);
4223
4224     for(i = 0; i < MAX_PPS_COUNT; i++)
4225         av_freep(h->pps_buffers + i);
4226 }
4227
4228 av_cold int ff_h264_decode_end(AVCodecContext *avctx)
4229 {
4230     H264Context *h = avctx->priv_data;
4231     MpegEncContext *s = &h->s;
4232
4233     ff_h264_remove_all_refs(h);
4234     ff_h264_free_context(h);
4235
4236     MPV_common_end(s);
4237
4238 //    memset(h, 0, sizeof(H264Context));
4239
4240     return 0;
4241 }
4242
4243 static const AVProfile profiles[] = {
4244     { FF_PROFILE_H264_BASELINE,             "Baseline"              },
4245     { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline"  },
4246     { FF_PROFILE_H264_MAIN,                 "Main"                  },
4247     { FF_PROFILE_H264_EXTENDED,             "Extended"              },
4248     { FF_PROFILE_H264_HIGH,                 "High"                  },
4249     { FF_PROFILE_H264_HIGH_10,              "High 10"               },
4250     { FF_PROFILE_H264_HIGH_10_INTRA,        "High 10 Intra"         },
4251     { FF_PROFILE_H264_HIGH_422,             "High 4:2:2"            },
4252     { FF_PROFILE_H264_HIGH_422_INTRA,       "High 4:2:2 Intra"      },
4253     { FF_PROFILE_H264_HIGH_444,             "High 4:4:4"            },
4254     { FF_PROFILE_H264_HIGH_444_PREDICTIVE,  "High 4:4:4 Predictive" },
4255     { FF_PROFILE_H264_HIGH_444_INTRA,       "High 4:4:4 Intra"      },
4256     { FF_PROFILE_H264_CAVLC_444,            "CAVLC 4:4:4"           },
4257     { FF_PROFILE_UNKNOWN },
4258 };
4259
4260 static const AVOption h264_options[] = {
4261     {"is_avc", "is avc", offsetof(H264Context, is_avc), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 1, 0},
4262     {"nal_length_size", "nal_length_size", offsetof(H264Context, nal_length_size), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 4, 0},
4263     {NULL}
4264 };
4265
4266 static const AVClass h264_class = {
4267     "H264 Decoder",
4268     av_default_item_name,
4269     h264_options,
4270     LIBAVUTIL_VERSION_INT,
4271 };
4272
4273 static const AVClass h264_vdpau_class = {
4274     "H264 VDPAU Decoder",
4275     av_default_item_name,
4276     h264_options,
4277     LIBAVUTIL_VERSION_INT,
4278 };
4279
4280 AVCodec ff_h264_decoder = {
4281     .name           = "h264",
4282     .type           = AVMEDIA_TYPE_VIDEO,
4283     .id             = CODEC_ID_H264,
4284     .priv_data_size = sizeof(H264Context),
4285     .init           = ff_h264_decode_init,
4286     .close          = ff_h264_decode_end,
4287     .decode         = decode_frame,
4288     .capabilities   = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 | CODEC_CAP_DELAY |
4289                       CODEC_CAP_SLICE_THREADS | CODEC_CAP_FRAME_THREADS,
4290     .flush= flush_dpb,
4291     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4292     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4293     .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4294     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4295     .priv_class     = &h264_class,
4296 };
4297
4298 #if CONFIG_H264_VDPAU_DECODER
4299 AVCodec ff_h264_vdpau_decoder = {
4300     .name           = "h264_vdpau",
4301     .type           = AVMEDIA_TYPE_VIDEO,
4302     .id             = CODEC_ID_H264,
4303     .priv_data_size = sizeof(H264Context),
4304     .init           = ff_h264_decode_init,
4305     .close          = ff_h264_decode_end,
4306     .decode         = decode_frame,
4307     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4308     .flush= flush_dpb,
4309     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4310     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_VDPAU_H264, PIX_FMT_NONE},
4311     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4312     .priv_class     = &h264_vdpau_class,
4313 };
4314 #endif